A Smart Normalization Method for Data Dictionary Names Based on Natural Language Generation

Abstract

This invention discloses an intelligent standardization method for data dictionary names based on natural language generation, belonging to the field of natural language processing technology. Combining deep reinforcement learning techniques, the method analyzes the contextual information of the input through a generative model, generates candidate terms, and compares them with predefined standardized terms. It employs three reward mechanisms: precision reward, recall reward, and context consistency reward. Through the combined effect of these three mechanisms, the generative model continuously optimizes its strategy and adjusts the generation process to achieve more accurate term generation. This method can flexibly adapt to terms in emerging fields and ensure term standardization through automated adjustment of the generation strategy. Especially in technical fields where terms are constantly changing and updated, it can effectively improve the efficiency and accuracy of term generation, thereby meeting rapidly changing needs.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and more specifically, to a method for intelligent standardization of data dictionary names based on natural language generation. Background Technology

[0002] With the rapid development of emerging technologies, especially breakthroughs in artificial intelligence, quantum computing, and blockchain, a plethora of related terms are emerging. Traditional terminology standardization methods largely rely on manual dictionaries, rule bases, or hand-designed dictionaries. These methods are inefficient when faced with emerging terms and struggle to meet rapidly changing needs. Due to the frequent updates and iterations of terminology, traditional manual standardization methods cannot adapt to the generation of new terms in a timely manner, resulting in delays and inefficiencies in the standardization process. Therefore, there is an urgent need to develop an intelligent and automated terminology standardization method that can quickly respond to the generation of new terms and accurately incorporate them into the standardization system to ensure terminology consistency and usability.

[0003] The existing technology has the following shortcomings:

[0004] Furthermore, traditional terminology standardization methods often fail to effectively address ambiguity and polysemy inherent in terminology generation. Standardization is particularly challenging for unfamiliar terms, lacking real-time feedback and adaptive capabilities, resulting in compromised accuracy and coverage. Existing technologies largely rely on static dictionaries or rule bases for terminology matching, which cannot quickly adapt to changing domains, especially when dealing with interdisciplinary or emerging technologies, where traditional methods often fall short of high-precision requirements. Therefore, there is an urgent need for a method that can automatically adapt, rapidly identify, and standardize new terms through intelligent learning, especially in complex and ever-changing technological environments, ensuring the accuracy and consistency of terminology generation.

[0005] To address the above problems, this invention proposes a solution. Summary of the Invention

[0006] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method for intelligent standardization of data dictionary names based on natural language generation, in order to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A method for intelligent standardization of data dictionary names based on natural language generation, including the following steps;

[0009] Step S1: Analyze the input context information, use the policy network in the deep reinforcement learning model to generate candidate terms, compare the generated terms with the standardized terms, calculate the matching degree between the generated terms and the standardized terms using the precision reward, calculate the coverage of the generated terms with the context-related terms using the recall reward, calculate the consistency between the generated terms and the context using the context consistency reward, and adjust the generation strategy according to the three mechanisms of precision reward, recall reward and context consistency reward.

[0010] Step S2: Compare the generated terms with standardized terms, calculate the matching degree using cosine similarity or Jaccard similarity, calculate the precision reward, recall reward and context consistency reward based on the comparison results and adjust the reward strategy to cover relevant terms, and filter terms that meet the standardization requirements by comparing them with standard terms in the knowledge graph.

[0011] Step S3: Among the filtered terms, calculate the semantic similarity between the terms and the context, and use the context consistency reward mechanism to perform term matching checks. The context consistency reward is calculated by generating terms and their context. If a term fails the consistency check, regenerate the term and make adjustments.

[0012] Step S4: Automatically adjust the generative model's strategy through a feedback mechanism. Based on the feedback results of three metrics—precision reward, recall reward, and context consistency reward—adjust the generative model's strategy, correct erroneous terms, calculate the reward based on the matching degree after each term generation, and use feedback signals from reinforcement learning to adjust the strategy and optimize the generation results.

[0013] In a preferred embodiment, step S1 includes the following:

[0014] Deep reinforcement learning models include policy networks and value networks;

[0015] The policy network generates candidate terms based on the contextual information of the input state;

[0016] The value network performs a quality assessment on the terms generated by the policy network.

[0017] Accuracy reward, recall reward, and context consistency reward together constitute the reward signal of reinforcement learning. The reward signal guides the policy network and value network to update their internal parameters, thereby adjusting and optimizing the term generation strategy.

[0018] Deep reinforcement learning models are trained and optimized using the A3C or PPO algorithm framework;

[0019] The policy network, acting as an actor network, is responsible for generating actions to output terms;

[0020] The value network, as a Critic network, is responsible for evaluating the expected value of the terms generated in the current state.

[0021] Precision reward, recall reward, and context consistency reward are integrated into the reward function of the algorithm framework, and the parameters of the Actor network and Critic network are iteratively updated through the policy gradient method.

[0022] In a preferred embodiment, step S2 includes the following:

[0023] Terms that meet the standardization requirements are selected by comparing them with standard terms in the knowledge graph;

[0024] Retrieve from the knowledge graph a set of standard terms that are semantically associated with the current input context;

[0025] The semantic matching degree between the generated candidate terms and the standard term set obtained from the query is calculated.

[0026] The calculation results are judged based on the preset matching degree threshold, and the generated terms with a matching degree higher than the threshold are selected as terms that initially meet the standardization requirements.

[0027] The semantic matching degree is specifically calculated using cosine similarity or Jaccard similarity algorithms, which convert the generated terms and standard terms in the knowledge graph into vector representations or word set representations before calculation.

[0028] The preset matching threshold is set according to the strictness of the domain standardization requirements to control the accuracy of the screening process.

[0029] In a preferred embodiment, step S3 includes the following:

[0030] Calculate the semantic similarity between terms and their context;

[0031] The terms to be inspected and their original input contexts are converted into high-dimensional semantic vectors respectively;

[0032] The cosine similarity between the semantic vector of a term and the semantic vector of its context is calculated, and this similarity value is used as a quantitative indicator to measure the semantic fit between the two.

[0033] If the terminology fails the consistency check, regenerate the terminology and adjust it accordingly;

[0034] When the semantic similarity between a term and its context is lower than a preset consistency threshold, the term is determined to have failed the consistency check.

[0035] The generative model, based on the updated policy network parameters and current context information, re-executes the term generation process to produce new candidate terms to replace those that failed the check.

[0036] In a preferred embodiment, step S4 includes the following:

[0037] The feedback mechanism is a policy update loop based on reinforcement learning;

[0038] Based on the feedback results of the three metrics—precision reward, recall reward, and context consistency reward—the strategy of the generative model is adjusted. After each term generation and evaluation cycle, the weighted sum of the three reward metrics is calculated as the total reward for this cycle. The total reward is then used to calculate the policy gradient through the backpropagation algorithm, thereby updating the network parameters of the deep reinforcement learning model.

[0039] The method involves multiple rounds of training.

[0040] In each round of training, the generative model processes multiple different contextual inputs and generates corresponding terms, and then summarizes all the generated results by an evaluation metric.

[0041] The optimized generation results are achieved through multiple rounds of training iterations. After each round of training, the model's generation strategy is fine-tuned based on a comprehensive evaluation of accuracy, recall, and F1-Score.

[0042] The technical effects and advantages of the intelligent standardization method for data dictionary names based on natural language generation in this invention are as follows:

[0043] Reinforcement learning-based generative models can rapidly adapt to unknown terms. Through a triple mechanism of precision rewards, recall rewards, and contextual consistency rewards, they ensure a high degree of matching between generated and standardized terms, significantly improving generation accuracy and coverage. Precision rewards ensure consistency between generated and standardized terms, recall rewards encourage coverage of more relevant terms, and contextual consistency rewards ensure that terms fit their context, avoiding misuse. When dealing with polysemous and ambiguous terms, the system automatically adjusts its generation strategy through the contextual consistency reward mechanism to ensure semantic accuracy. Through feedback mechanisms, the generative model continuously self-corrects and optimizes, improving the accuracy of term generation, adapting to rapidly changing new terms, and providing effective standardization solutions for interdisciplinary and emerging fields. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the intelligent standardization method for data dictionary names based on natural language generation according to the present invention. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] Example: Please refer to Figure 1 As shown, this invention discloses an intelligent standardization method for data dictionary names based on natural language generation, including the following steps:

[0047] Step S1: Analyze the input context information, use the policy network in the deep reinforcement learning model to generate candidate terms, compare the generated terms with the standardized terms, calculate the matching degree between the generated terms and the standardized terms using the precision reward, calculate the coverage of the generated terms with the context-related terms using the recall reward, calculate the consistency between the generated terms and the context using the context consistency reward, and adjust the generation strategy according to the three mechanisms of precision reward, recall reward and context consistency reward.

[0048] Step S2: Compare the generated terms with standardized terms, calculate the matching degree using cosine similarity or Jaccard similarity, calculate the precision reward, recall reward and context consistency reward based on the comparison results and adjust the reward strategy to cover relevant terms, and filter terms that meet the standardization requirements by comparing them with standard terms in the knowledge graph.

[0049] Step S3: Among the filtered terms, calculate the semantic similarity between the terms and the context, and use the context consistency reward mechanism to perform term matching checks. The context consistency reward is calculated by generating terms and their context. If a term fails the consistency check, regenerate the term and make adjustments.

[0050] Step S4: Automatically adjust the generative model's strategy through a feedback mechanism. Based on the feedback results of three metrics—precision reward, recall reward, and context consistency reward—adjust the generative model's strategy, correct erroneous terms, calculate the reward based on the matching degree after each term generation, and use feedback signals from reinforcement learning to adjust the strategy and optimize the generation results.

[0051] In step S1, the input context information is analyzed, and candidate terms are generated using the policy network in the deep reinforcement learning model. The generated terms are compared with standardized terms. Precision rewards are used to calculate the matching degree between the generated terms and standardized terms, recall rewards are used to calculate the coverage of the generated terms with context-related terms, and context consistency rewards are used to calculate the consistency between the generated terms and the context. The generation strategy is then adjusted based on these three mechanisms: precision reward, recall reward, and context consistency reward. Specifically, the following steps are included:

[0052] The specific design of the policy network and value network in the deep reinforcement learning model is as follows: The policy network adopts a multi-layer convolutional neural network structure, containing 4 convolutional layers, and the activation function is ReLU;

[0053] The input is a semantic vector of the context, and the output is the probability distribution of the generated terms. The value network adopts a fully connected layer structure, which contains 3 neural networks with ReLU activation function. The output is the estimated value of term quality.

[0054] In this invention, semantic vector conversion uses the BERT model, employing 768-dimensional vector representations to generate terms and standardize terms. Before being converted into semantic vectors, all terms are first encoded using the BERT model to obtain word embedding representations. Semantic vectorization processing is performed on generated terms, contextual information, and standardized terms respectively, providing a foundation for subsequent similarity calculations.

[0055] The context information consists of the main parts of the input text, including the title, keywords, and context sentences. The context extraction method uses semantic analysis based on the BERT model to extract the core information in the context and generate semantic vectors.

[0056] The multi-agent parallel training mechanism of the A3C algorithm uses 8 agents for parallel training. By sharing the parameters of the global network and the local network, it adopts an asynchronous gradient update method, and uses a bidirectional synchronization method for data synchronization to ensure that the parameter updates between the global network and the local network are consistent.

[0057] The PPO algorithm uses a pruning factor of 0.2 and the generalized dominance estimation method is used to calculate the dominance function.

[0058] With the continuous development of emerging fields, such as artificial intelligence, quantum computing, and blockchain, new terms are constantly emerging. Traditional terminology standardization methods based on dictionaries and rules cannot meet the needs of dynamic changes. By optimizing the generative model through reinforcement learning, it can quickly adapt to and standardize the generation of new terms that have not been seen before.

[0059] The generative model is trained using deep reinforcement learning techniques and employs a policy network and a value network to jointly optimize the generation process. In this process, the generative model acts as an agent, maximizing the objective function through interaction with the environment. The generative model first analyzes the input context and generates generative terms. Generative terms refer to terms generated by the generative model based on the given context. These terms may not yet be defined in a standardized dictionary or data dictionary. Therefore, the generated terms will be compared with standardized terms. Standardized terms refer to terms that have been reviewed by experts, processed by knowledge graphs or standardization processes, and conform to the standards in the domain. The comparison between generated terms and standardized terms is used to ensure that the generated terms meet the standardization requirements. During the comparison process, the accuracy and applicability of the generated terms will be verified, and the generative model will be adjusted according to the comparison results to ensure the consistency and standardization of the final generated terms with the standardized terms.

[0060] Here, generated terms refer to the candidate data dictionary names output by the generative model in the current input context. These terms are generated by the model based on the context and may not yet be defined in the existing standard data dictionary. Standardized terms refer to the canonical names registered in the pre-built standard data dictionary or domain knowledge graph, or the canonical names pointed to by synonym mapping relationships. Subsequent comparisons and reward calculations are based on generated terms and standardized terms. The comparisons ensure that the generated terms meet the standardization requirements, and the generative model is optimized based on the comparison results.

[0061] According to the reward mechanism, the generative model receives a positive reward when it successfully generates terms that meet the standardized terminology requirements, and a negative penalty when the generated terms fail to meet the standardized terminology requirements. This reward and penalty-based learning approach promotes the gradual improvement of the generative model, enabling it to more accurately identify and standardize emerging terms. Whenever the model successfully generates terms that meet the standardized requirements, the reward mechanism will provide positive feedback, prompting the model to adjust towards a more accurate generation strategy; conversely, when the generated terms do not meet the standardized requirements, negative feedback will prompt the model to optimize, thereby adapting to more new terms and application scenarios during continuous training and adjustment.

[0062] To ensure that the generative model can efficiently identify and generate standardized terms in rapidly changing domains, this step designs a system model based on reinforcement learning, specifically using a deep reinforcement learning framework. This framework uses deep neural networks to construct a policy network and a value network, and optimizes decisions through the interaction between the agent and the environment. The policy network is responsible for generating corresponding terms based on the current state, while the value network evaluates the quality of the generated terms, helping the agent optimize the decision-making process to ensure that the generated terms meet the standardization requirements.

[0063] The constructed model mainly consists of a policy network and a value network, which work together to optimize the decision-making of generative AI in the term generation process. The policy network generates corresponding terms based on the current state, while the value network evaluates the quality of the generated terms and influences subsequent decisions based on the evaluation results. Through a reinforcement learning framework, the model continuously adjusts its policy during the generation process to improve the standardization of the generated terms. Specifically, the actor-critic method of A3C or PPO is used to optimize the generation process, where the value network is used to estimate the expected quality of the generated terms and guide the policy network to adjust parameters and update decisions.

[0064] More specifically, both the policy network and the value network are modeled using deep neural networks. The input of the network is the state representation of the current generation environment, i.e., contextual information, and the output is the generated term or the quality evaluation value of the term. Through repeated training, the model learns how to generate terms that meet the standardization requirements in different contexts by adjusting the parameters in the policy network and the value network. During the training process, the generated terms are compared with the standardized terms, and the model's generation strategy is optimized based on the comparison results to ensure that the model can adapt to the ever-changing domain and emerging terms.

[0065] To ensure that the generative model can react quickly to emerging terms and generate terms that meet standardization requirements, this step uses three key metrics—precision reward, recall reward, and context consistency reward—to guide the optimization of the model.

[0066] The accuracy reward is used to measure the matching degree of the terms generated by the model. The accuracy reward can be calculated using the following formula: ;in, The accuracy reward measures the degree of matching between generated terms and standardized terms. These are generated terms, that is, the terms output by the generative model in a given context. These are standardized terms, obtained through domain knowledge graphs or standard dictionaries.

[0067] The matching degree between generated terms and standardized terms is calculated by using cosine similarity or Jaccard similarity. High similarity will give the model a higher reward, thereby promoting the generation of terms that are more in line with the standard. The accuracy reward ensures that the generated terms are highly consistent with the standardized terms, prompting the model to maintain high accuracy during the generation process. Through this reward mechanism, the generative model learns how to generate standardized terms, especially when faced with terms in emerging fields, it can quickly adapt and make standardized judgments.

[0068] Standardized terms are derived from recognized standard dictionaries or knowledge graphs in the field. Generated terms are candidate terms generated by deep learning models in specific contexts and have not yet been formally standardized. The scope of generated terms includes, but is not limited to, terms in existing knowledge graphs. The model can propose new terms during the generation process and make standardization judgments.

[0069] Besides accuracy rewards, recall rewards also play a crucial role. Recall rewards encourage models to identify and generate terms relevant to the current context. The more relevant terms a model can identify, even if they don't fully conform to standardized terms, the more it can improve its performance through recall rewards. The calculation of recall rewards is as follows: ;in, This represents the recall reward, which measures the overlap between the terms generated by the model and the set of terms relevant to the context. It is a term that is generated. It is a set of terms relevant to the current context, representing the terms that the model should recognize and generate;

[0070] Recall rewards help models expand the range of terms they generate, ensuring that the generated terms not only meet standardization requirements but also cover more context-related terms. Through recall rewards, models can improve the recognition rate of relevant terms and play a positive role in the standardization of emerging terms.

[0071] Contextual consistency reward measures the fit between generated terms and their context, preventing terms from appearing in inappropriate contexts. Contextual consistency is crucial for the standardization process; generated terms must remain consistent with their context to ensure accuracy in practical applications. The contextual consistency reward is calculated as follows: ;in, This represents a context consistency reward, measuring the consistency between generated terms and their context. It refers to the context in which the term is generated, that is, the linguistic context in which the term is generated. It refers to the context in which standardized terms are used;

[0072] By calculating the similarity between the generated terms and the context of the standardized terms, it can be ensured that the generated terms are semantically and structurally consistent with the context. If the generated terms fail to fit the context correctly, the model will be penalized. The context consistency reward improves the quality of the standardized terms by ensuring that the generated terms fit the context, and avoids errors and inconsistencies in the term generation process.

[0073] Through the combined effect of precision rewards, recall rewards, and context consistency rewards, the model can maintain high precision and high recall during the generation of new terms, and ensure that the generated terms conform to the context and standardization requirements. These reward mechanisms work together to drive the generative model to continuously optimize its generation strategy, enabling it to standardize quickly and accurately when faced with emerging terms.

[0074] Guided by reinforcement learning, the training objective of the generative model is to optimize the model's generation strategy through a reward mechanism. After each term is generated and a reward is calculated, the generative model will adjust its internal parameters based on the reward feedback, so that the model can generate terms that are more in line with standardization requirements in future generation processes.

[0075] To ensure that the model can quickly adapt to changes in new terms, reinforcement learning algorithms A3C and PPO were used during training. These algorithms enable parallel training through multiple agents, allowing the model to optimize and adapt to different terminology generation tasks more quickly.

[0076] The A3C algorithm trains multiple agents in parallel, enabling training and optimization in different environments and improving the model's generalization ability. The PPO algorithm, on the other hand, avoids the over-update problem in reinforcement learning by optimizing the update of the strategy, ensuring the stability of the model during training. Using these algorithms, the generative model can effectively learn how to generate new terms that meet standardization requirements and gradually optimize the generation strategy based on the reward mechanism.

[0077] To ensure that the generative model can quickly adapt to changes in new terms, data augmentation techniques are used during training. Through data augmentation, the model can be exposed to more types of new terms. The diversity of training data enables the generative model to generate terms that meet standardization requirements in different contexts. The design not only includes existing standardized terms in the field, but also terms and concepts that are constantly emerging in emerging fields. Through continuous training and data iteration, the generative model can quickly identify and generate standardized terms related to these new terms.

[0078] Evaluation metrics during model training include accuracy, recall, and F1-Score. Accuracy measures the degree of matching between generated terms and standardized terms; recall measures whether the generated terms cover all relevant terms; and F1-Score takes into account both accuracy and recall to evaluate the overall performance of the model. Through these evaluation metrics, the model training process can be continuously optimized to ensure that it can maintain high accuracy and high efficiency when generating new terms.

[0079] Through multiple rounds of training and optimization, the generative model can quickly identify and generate terms that meet standardization requirements when faced with emerging terms. The combined effects of reinforcement learning, reward mechanisms, parallel training, and data augmentation techniques enable the generative model to be efficient and accurate in handling emerging terms. This process not only ensures that new terms can be quickly incorporated into the standardization system, but also promotes the generative model to maintain high adaptability and performance in the ever-changing technological field.

[0080] The knowledge graph consists of standardized terms and related concepts within the domain. The graph structure is based on an entity relationship graph, where each term is an entity, and the relationships between each pair of terms are connected by edges such as synonyms, near-synonyms, and hierarchical relationships. The query algorithm uses a semantic similarity-based query method to calculate the similarity between generated terms and standard terms in the knowledge graph, and selects the most matching terms. The knowledge graph update mechanism adopts automated learning, which automatically expands and updates the semantics through continuous input of new terms.

[0081] By introducing reinforcement learning and a refined reward mechanism, a highly efficient generative AI model was designed. This model can quickly adapt to emerging terms and standardize them by combining precision rewards, recall rewards, and context consistency rewards. This ensures that the generative model can generate terms that meet standardization requirements and adapt to the ever-changing terminology environment. Through multiple rounds of training, the generative model continuously optimizes its generation strategy, ultimately enabling it to accurately identify new terms and standardize them, thus guaranteeing high quality and efficiency in terminology generation.

[0082] In step S2, the generated terms are compared with standardized terms, and the matching degree is calculated using cosine similarity or Jaccard similarity. Based on the comparison results, precision reward, recall reward, and context consistency reward are calculated, and reward adjustment strategies are implemented to cover relevant terms. By comparing with standard terms in the knowledge graph, terms that meet the standardization requirements are selected. Specific content includes:

[0083] During the generation process, the goal of the generative AI model is to accurately generate terms that meet the standardization requirements. It relies on the domain knowledge graph to provide background knowledge support for the generative model. As an auxiliary verification mechanism, the knowledge graph provides a powerful term matching system when generating terms. It can help the model ensure that the generated terms meet the domain standards through comparison and reasoning, and avoid situations where the generated terms cannot match known terms or cause ambiguity.

[0084] When generating new terms, the model automatically queries the knowledge graph related to the current context, performs term reasoning, and compares the generated terms with the standardized terms in the knowledge graph. If there is consistency, the generated terms pass the verification; if there is a discrepancy, the model will correct and optimize according to the synonym mapping rules in the knowledge graph.

[0085] With the introduction of knowledge graphs, generative AI will enhance its ability to filter terms in term generation. This filtering process is based on contextual semantic analysis, which compares the generated terms with standard terms in the knowledge graph to determine whether the generated terms conform to the context and domain requirements.

[0086] In this process, generative AI uses Transformer-based models, such as BERT or GPT, to deeply analyze the context of the input and extract semantic information related to the terms. By comparing the semantic matching degree between the vocabulary and syntactic structure in the current context and the generated terms, the model can decide whether the generated terms meet the expected standards. If the terms do not match the context, they are excluded through a semantic similarity mechanism to ensure that the generated terms meet the standardization requirements.

[0087] Regarding threshold setting, if the semantic similarity between the generated term and the context is lower than the set threshold, such as a cosine similarity of 0.7, the term will be automatically excluded, and a new term will be generated or a term that meets the standard will be selected from the candidate terms.

[0088] If the generated terms are semantically inconsistent and fail the contextual consistency check, they are deemed invalid terms and need to be regenerated to meet the standardization requirements.

[0089] The semantic similarity between generated terms and their context can be calculated using cosine similarity. ;in, To generate vector representations of terms. It is a vector representation of the context, and the calculation results are used to determine the consistency between the term and the context;

[0090] The generated candidate terms will enter the screening mechanism. This process not only conducts preliminary screening through context analysis, but also conducts further filtering by combining semantic matching algorithms. By using models such as Jaccard similarity or cosine similarity, the generated terms are compared with the terms in the knowledge graph to ensure that the semantics of the generated terms are consistent with the standard terms.

[0091] Precision and recall, two important metrics, will be used to measure the effectiveness of term selection. Precision calculates the degree of matching between generated terms and standardized terms, while recall ensures that generated terms can cover a sufficient number of relevant terms, especially for terms in emerging fields that have not yet been clearly standardized.

[0092] The formula for calculating precision can be expressed as: The recall rate can be calculated using the following formula: ;in, Represents the generation of term sets A standardized set of terms;

[0093] When further optimizing the filtered terms, the generative AI will adopt a context consistency check mechanism. The context consistency reward is used to measure the fit between the generated terms and their context, ensuring that the terms do not appear in inappropriate contexts.

[0094] This reward mechanism ensures that terms not only meet semantic requirements, but also fit their context. For example, if the terms generated by the model do not match the technical content when describing the background of a new technology, the generation strategy will be adjusted through negative feedback.

[0095] The context consistency threshold is The cosine similarity is set to, for example, 0.8. If the similarity of the generated terms does not reach this threshold, the generation process will backtrack and make adjustments until it meets the standard.

[0096] The context consistency reward can be quantified by calculating the similarity between the generated term and its context, as shown in the following formula: ;in, To generate terms, As a semantic representation of the context, if the degree of fit between the generated term and the context is lower than the threshold, it is judged as an unqualified term and needs to be regenerated;

[0097] Based on the integrated design of deep reinforcement learning and knowledge graph, the generation process is optimized by using A3C or PPO algorithms. A3C adopts multi-agent parallel training, which improves training efficiency and model generalization ability, while PPO avoids the over-update problem in reinforcement learning by restricting policy updates.

[0098] The policy network takes contextual information as input and generates terms as output. It optimizes the policy through backpropagation and adjusts the generation policy accordingly.

[0099] The value network is used to evaluate the quality of generated terms and guide the optimization of the policy network. The model optimizes the generation policy by estimating the standardization fitness of the generated terms.

[0100] After multiple rounds of training and optimization, the generated terms will work together based on precision rewards, recall rewards and context consistency rewards to ensure that the generated terms meet the standardization requirements. The reward mechanism after each generation will drive the model to continuously optimize its generation strategy, enabling it to quickly adapt to and accurately standardize emerging terms.

[0101] To ensure that the generative model can operate stably in a constantly changing terminology environment, data augmentation techniques were used during training to enable the model to learn on various types of terminology datasets, thereby improving its adaptability in emerging fields.

[0102] The training process of the model uses metrics such as precision, recall, and F1-score. Precision measures the matching degree of generated terms, recall measures the coverage of generated terms, and F1-score combines precision and recall to evaluate the overall performance of the model.

[0103] By combining generative AI and knowledge graphs, and by optimizing the screening mechanism and introducing a feedback mechanism, the standardization accuracy of new terms is effectively improved. Each step is interconnected, ensuring accurate conversion from generated terms to standardized terms. In addition, through reinforcement learning and reward mechanisms, the model can quickly adapt to and maintain efficiency and accuracy when faced with constantly changing new terms.

[0104] In step S3, among the filtered terms, the semantic similarity between the terms and their context is calculated. A context consistency reward mechanism is used to perform term matching checks. The context consistency reward is calculated by determining the fit between the generated term and its context. If a term fails the consistency check, the term is regenerated and adjusted. Specifically, this includes:

[0105] Based on step S2, the re-evaluation of the screened terms is further refined, and special cases are addressed to ensure that the generated terms meet more stringent standardization requirements.

[0106] Polysemous words and ambiguous terms are common problems in the process of terminology standardization, especially in emerging fields. Generative models face uncertainty and ambiguity in the meaning of terms. In order to accurately determine the meaning of terms, it is necessary to rely on contextual information for weighted calculation to ensure that polysemous words and ambiguous terms can be correctly interpreted according to the specific context.

[0107] The semantic accuracy of generated terms is evaluated using contextual similarity, and the formula can be expressed as: ;in, To generate terms, Calculate the accuracy of the term in multiple contexts, given information from different contexts;

[0108] The context-weighted mechanism ensures the correct interpretation of polysemous terms by performing multi-dimensional semantic analysis on the terms, avoiding the influence of ambiguity. The calculation of context similarity guarantees the accuracy of the terms. Through this mechanism, the model can make the most suitable standardized judgment when encountering polysemous words.

[0109] The handling of terms with unclear context is very important, especially when a term cannot be found in the domain knowledge graph or the context is not entirely clear. To solve this problem, it is necessary to rely on external corpora and large model reasoning to determine the most likely meaning of the term, thereby generating terms that meet the standardization requirements.

[0110] For terms with unclear context, standardized inference scores are used for judgment, and the formula can be expressed as: ;in, To generate terms, Let n be the number of candidate standard terms, and calculate the average similarity of these terms.

[0111] By combining reasoning mechanisms with external corpora, the problem of unclear terminology context can be solved, which is particularly effective in the generation of terminology in emerging fields and interdisciplinary fields. This step ensures accurate reasoning of terminology and avoids erroneous standardization caused by unclear context.

[0112] In certain specialized fields, such as medicine and finance, the standardization requirements for terminology are particularly stringent. These fields require a high degree of specialization and precision. The generated terms must strictly match the standard terms in the domain knowledge graph to ensure their professionalism and compliance with industry standards. The domain knowledge graph plays a crucial role in this process. The domain adaptation scoring formula calculates the semantic matching degree between the generated terms and the domain's standard terms. ;in, To generate terms, For standard terms in the domain knowledge graph, calculate the maximum value to select the most appropriate term;

[0113] The assistance of domain knowledge graphs enables the generated terms to be evaluated in an industry-specific context. By matching with the domain knowledge graph, it can be ensured that the generated terms conform to industry standards and avoid the appearance of non-standard terms.

[0114] Feedback mechanisms are crucial for optimizing the generative model. After terms are generated, the model continuously optimizes its generation strategy through an automatic correction mechanism. If a generated term fails validation, the feedback mechanism prompts the model to automatically adjust its parameters, thereby improving the model's accuracy in term generation. Through iterative optimization, the model can continuously adapt to new terms and application scenarios. The optimization of the feedback mechanism is achieved by adjusting the model parameters, and the update formula is as follows: ;in, For model parameters, For the feedback loss function, Let m be the learning rate, and m be the number of feedback updates.

[0115] Through backpropagation and feedback mechanisms, the model can correct its generation strategy in each update, thereby continuously improving the accuracy of generated terms. The automatic adjustment forms a virtuous cycle, ensuring that the model gradually learns how to generate terms that conform to the standards.

[0116] In each round of training, the model needs to optimize its generation strategy based on evaluation metrics such as precision, recall, and F1-Score. Precision and recall are important metrics for measuring the effectiveness of term standardization. Through these two metrics, the model can know in which aspects it performs well and in which aspects it still needs to improve.

[0117] The model is evaluated using precision, recall, and F1-Score. These evaluation metrics provide a clear direction for optimization. Based on feedback from precision and recall, the model can continuously adjust its generation strategy to ensure that the generated terms not only meet standardization requirements but also cover more relevant terms.

[0118] All term generation and evaluation need to be standardized as the ultimate goal. Through multiple rounds of training and optimization, the model can react quickly when faced with emerging terms and generate terms that meet the standardization requirements. Each optimization and adjustment enables the model to standardize more accurately.

[0119] During the optimization process, the model continuously adjusts the precision reward, recall reward, and context consistency reward to improve its standardization effect, ultimately generating standardized terms. ;in, As the final reward, A reward is calculated based on the degree of matching between the generated terms and the target terms;

[0120] By combining multiple reward mechanisms, the model can balance precision, recall, and contextual consistency, ultimately generating terms that meet standardization requirements.

[0121] By introducing a multi-dimensional re-evaluation mechanism and methods to address special cases, such as polysemous words and ambiguous contexts, the model ensures the standardization accuracy of generated terms in different contexts. With the assistance of reinforcement learning, context analysis, and domain knowledge graphs, the model continuously adjusts its generation strategy to achieve efficient and accurate terminology standardization. Through reward and feedback mechanisms, the model can quickly adapt to and maintain consistency in the ever-changing terminology environment.

[0122] In step S4, the strategy of the generative model is automatically adjusted through a feedback mechanism. Based on the feedback results of three metrics—precision reward, recall reward, and context consistency reward—the strategy of the generative model is adjusted to correct generated erroneous terms. After each term generation, a reward is calculated based on the matching degree. The strategy is adjusted using feedback signals from reinforcement learning to optimize the generation results. Specific details include:

[0123] In the process of automated term generation, machine learning models may make some errors, especially when the terms do not have sufficient context or domain knowledge support. Therefore, an automatic correction mechanism is the key to improving the standardization accuracy of the model. Through reinforcement learning algorithms, the model can automatically correct potential errors in the generated terms based on historical error feedback and training data, thereby improving the standardization quality of terms.

[0124] After generating terms, if the generated terms do not meet the standardization requirements, the term will be automatically recorded and adjusted. It will self-correct based on the previously generated erroneous terms and update the parameters of the generation model so that it can avoid the same error in the next round of generation. This process is completed through the interaction of the value network and policy network in reinforcement learning. Through the backpropagation algorithm, the automatic correction mechanism continuously optimizes the model's decision-making, thereby reducing the probability of erroneous generation.

[0125] The automatic correction mechanism relies on the combination of the value network and the policy network in reinforcement learning. It adjusts the model's generation policy through reward signals, as shown in the following formula: ;in, To generate the parameters of the model, For the automatically corrected loss function, Let m be the learning rate, and m be the number of correction updates.

[0126] The automatic correction mechanism automatically optimizes the model based on feedback from previously generated errors. Through backpropagation of erroneous terms, the model gradually learns how to avoid similar errors, thereby ensuring that the terms it generates are more in line with standardization requirements.

[0127] Within the framework of reinforcement learning, the model generation process not only relies on automatic correction, but also needs to adjust the generation strategy through adaptive optimization. The adaptive optimization mechanism dynamically adjusts the generation model's strategy through algorithms, enabling it to adapt to new terminology generation tasks and quickly adjust its strategy to improve the standardization accuracy of terms. In particular, when facing terms in emerging fields, the model needs to have flexible adaptability.

[0128] After generating terms, rewards are calculated based on the degree of matching with standardized terms, and the model is optimized through a reinforcement learning reward mechanism. During the optimization process, the parameters of its policy network and value network are continuously adjusted to drive the model to learn how to generate standardized terms.

[0129] The adaptive optimization process adjusts the parameters of the generated model using the following formula: ;in, As the final reward, For accuracy rewards, As a recall reward, As a reward for context consistency, A penalty mechanism for generating errors; through these reward signals, the model can adaptively optimize its generation strategy.

[0130] The adaptive optimization mechanism combines rewards from multiple dimensions such as precision, recall, and context consistency. By adjusting the strategy of the generative model, it enables it to quickly adapt to and optimize the generation process when dealing with new terms, thereby reducing errors.

[0131] In each round of training, the model needs to continuously optimize its generation strategy based on evaluation metrics such as precision, recall, and F1-score. Precision and recall, as evaluation standards, help the model understand where it performs well and where it needs improvement. When dealing with complex terminology generation tasks, multiple rounds of training can ensure that the generation model gradually approaches perfect standardized terminology generation.

[0132] After each training round, the generated terms are compared with standardized terms, and precision and recall are calculated. The F1-Score is used to comprehensively evaluate the model's generation ability. The model adjusts its generation strategy based on these evaluation results, so that errors in the term generation process are gradually reduced.

[0133] During multiple rounds of training, the model's parameters can be updated using the following formula: ;in, To generate the parameters of the model, For the final loss function, Where n is the learning rate and n is the number of training rounds;

[0134] After each training round, the model adjusts its generation strategy based on the evaluation results of precision and recall. Through multiple rounds of training and incremental optimization, the model can gradually improve the standardization quality of generated terms and quickly adapt to terms in emerging fields.

[0135] In the process of generating terms, reward mechanisms, such as precision rewards, recall rewards, and context consistency rewards, can help the model improve the quality of term generation. However, reward mechanisms alone are not enough to ensure high quality of term generation. By combining adaptive optimization and incremental correction mechanisms, the model can further optimize the generation process and improve the standardization quality of generated terms.

[0136] The reward mechanism guides the generative model to optimize the generation strategy, while the adaptive optimization and incremental correction mechanism continuously updates the generation strategy, enabling it to avoid errors in each generation and gradually improve the quality of terminology standardization.

[0137] The final calculation formula combining the reward mechanism and the correction mechanism is as follows: ;in, As the final reward, The reward is calculated based on the matching degree between the generated terms and the target terms;

[0138] By combining reward and correction mechanisms, the model's generation strategy can not only optimize accuracy and recall, but also improve contextual consistency and domain adaptability. The resulting terms meet higher standardization requirements and can adapt to the ever-changing terminology environment.

[0139] After multiple rounds of training and optimization, the model needs to be evaluated using metrics such as precision, recall, and F1-Score. Based on these evaluation results, the model can adjust its generation strategy so that terminology generation not only meets standardization requirements but also covers more relevant terms, especially in rapidly generating standard-compliant terms in emerging fields.

[0140] In the final optimization phase, the model adjusts its generation strategy based on the evaluation results to ensure that the accuracy, recall, and contextual consistency of term generation all meet the expected goals. Through continuous training and adjustments, the model can stably generate terms that meet standardization requirements in constantly changing domains.

[0141] The final optimized evaluation metrics are as follows: Using F1-Score as the final evaluation metric, the model must ensure not only high accuracy of generated terms, but also high recall. Taking both aspects into account, the model ensures that the final generated terms achieve optimal standardization.

[0142] By combining automatic correction and adaptive optimization mechanisms, the model can continuously optimize its strategy in each generation round, improving the quality of generated terms. Multiple rounds of training and incremental optimization ensure that the model can quickly adapt when generating new terms and generate terms that meet standardization requirements. Through continuous optimization of reward and feedback mechanisms, the generation model can quickly adapt to and maintain consistency in the ever-changing terminology environment, ultimately generating accurate and standardized terms.

[0143] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0144] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0145] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and inventive constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0146] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0147] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0148] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A natural language generation-based data dictionary name intelligent standardization method, characterized in that, Includes steps; Step S1: Analyze the input context information, use the policy network in the deep reinforcement learning model to generate candidate terms, compare the generated terms with the standardized terms, calculate the matching degree between the generated terms and the standardized terms using the precision reward, calculate the coverage of the generated terms with the context-related terms using the recall reward, calculate the consistency between the generated terms and the context using the context consistency reward, and adjust the generation strategy according to the three mechanisms of precision reward, recall reward and context consistency reward. Step S2: Compare the generated terms with standardized terms, calculate the matching degree using cosine similarity or Jaccard similarity, calculate the precision reward, recall reward and context consistency reward based on the comparison results and adjust the reward strategy to cover relevant terms, and filter terms that meet the standardization requirements by comparing them with standard terms in the knowledge graph. Step S3: Among the filtered terms, calculate the semantic similarity between the terms and the context, and use the context consistency reward mechanism to perform term matching checks. The context consistency reward is calculated by generating terms and their context. If a term fails the consistency check, regenerate the term and make adjustments. Step S4: Automatically adjust the generative model's strategy through a feedback mechanism. Based on the feedback results of three metrics—precision reward, recall reward, and context consistency reward—adjust the generative model's strategy, correct erroneous terms, calculate the reward based on the matching degree after each term generation, and use feedback signals from reinforcement learning to adjust the strategy and optimize the generation results. 2.The natural language generation based data dictionary name intelligent standardization method of claim 1, wherein, Deep reinforcement learning models include policy networks and value networks; The policy network generates candidate terms based on the contextual information of the input state; The value network performs a quality assessment on the terms generated by the policy network. Accuracy rewards, recall rewards, and context consistency rewards together constitute the reward signal in reinforcement learning. The reward signal guides the policy network and value network to update their internal parameters, thereby adjusting and optimizing the term generation strategy. 3.The natural language based data dictionary name intelligent standardization method of claim 2, wherein, Deep reinforcement learning models are trained and optimized using the A3C or PPO algorithm framework; The policy network, acting as an actor network, is responsible for generating actions to output terms; The value network, as a Critic network, is responsible for evaluating the expected value of the terms generated in the current state. Precision reward, recall reward, and context consistency reward are integrated into the reward function of the algorithm framework, and the parameters of the Actor network and Critic network are iteratively updated through the policy gradient method.

4. The intelligent standardization method for data dictionary names based on natural language generation according to claim 1, characterized in that, Terms that meet the standardization requirements are selected by comparing them with standard terms in the knowledge graph; Retrieve from the knowledge graph a set of standard terms that are semantically associated with the current input context; The semantic matching degree between the generated candidate terms and the standard term set obtained from the query is calculated. The calculation results are judged based on a preset matching degree threshold, and the generated terms with a matching degree higher than the threshold are selected as terms that initially meet the standardization requirements.

5. The intelligent standardization method for data dictionary names based on natural language generation according to claim 4, characterized in that, The semantic matching degree is specifically calculated using cosine similarity or Jaccard similarity algorithms, which convert the generated terms and standard terms in the knowledge graph into vector representations or word set representations before calculation. The preset matching threshold is set according to the strictness of the domain standardization requirements. In the similarity calculation between terms and standardized terms, if the matching degree is higher than the set threshold, it is determined to meet the standardization requirements.

6. The intelligent standardization method for data dictionary names based on natural language generation according to claim 1, characterized in that, Calculate the semantic similarity between terms and their context; The terms to be inspected and their original input contexts are converted into high-dimensional semantic vectors respectively; The cosine similarity between the semantic vector of a term and the semantic vector of its context is calculated, and this similarity value is used as a quantitative indicator to measure the semantic fit between the two.

7. The intelligent standardization method for data dictionary names based on natural language generation according to claim 6, characterized in that, If the terminology fails the consistency check, regenerate the terminology and adjust it accordingly; The semantic similarity measure between terms and their context is used. When the semantic similarity of the generated terms is lower than the consistency threshold, the terms are deemed to have failed the consistency check. The generative model, based on the updated policy network parameters and current context information, re-executes the term generation process to produce new candidate terms to replace those that failed the check.

8. The intelligent standardization method for data dictionary names based on natural language generation according to claim 1, characterized in that, The feedback mechanism is a policy update loop based on reinforcement learning; Based on the feedback results of the three metrics—precision reward, recall reward, and context consistency reward—the strategy of the generative model is adjusted. After each term generation and evaluation cycle, the weighted sum of the three reward metrics is calculated as the total reward for this cycle. The total reward is then used to calculate the policy gradient through the backpropagation algorithm, thereby updating the network parameters of the deep reinforcement learning model.

9. The intelligent standardization method for data dictionary names based on natural language generation according to claim 8, characterized in that, The method involves multiple rounds of training. In each round of training, the generative model processes multiple different contextual inputs and generates corresponding terms, and then summarizes all the generated results by an evaluation metric. The optimized generation results are achieved through multiple rounds of training iterations. After each round of training, the model's generation strategy is fine-tuned based on a comprehensive evaluation of accuracy, recall, and F1-Score.

Images