Dense point identification method and device combining entropy guidance and large language model test enhancement

Abstract

The invention relates to a dense point recognition method and device combining entropy guidance and large language model test enhancement. The method comprises the following steps: firstly, through oversampling balance training data, constructing a classification model by using a lightweight pre-training model; dividing a to-be-recognized document into abstract and sentence sequences, and inputting the abstract and sentence sequences into the quantized large language model to generate an enhanced sample; and screening effective samples according to the entropy variation, and inputting the effective samples and the original sentences into a classification model. And processing a prediction result through a preset aggregation strategy, and finally outputting the security level of each sentence. The process efficiently balances data and enhances samples, and improves security classification recognition accuracy. By adopting the method, the problems of poor generalization performance of dense point recognition, uncontrollable enhanced sample quality, context semantic loss and insufficient confidential environment adaptability in a few-sample scene can be solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of natural language processing and security management, in particular to a secret point identification method and device combining entropy guidance and large language model test enhancement. BACKGROUND

[0002] Secret point identification is the core task of precise security management, which aims to split the secret document into independent "secret matters" and accurately determine the corresponding secret level, providing technical support for security review and information management. With the development of natural language processing technology, the technical route of secret point identification has evolved from traditional keyword matching and regular expression matching to text classification methods based on deep learning models such as TextCNN and BERT, significantly improving recognition accuracy and efficiency. In actual secret scenarios, model training is subject to strict security regulations, and secret data and model training environments must be physically isolated, which prevents the construction of large-scale labeled data sets and only allows for a small amount of labeled secret text data. Existing deep learning models rely on large-scale labeled data for training, and in the case of few samples, overfitting is likely to occur, and the generalization performance is significantly reduced, making it difficult to meet the accuracy requirements of secret point identification.

[0003] To solve the problem of data scarcity, data augmentation techniques are widely used in text classification tasks, and test-time augmentation (TTA) techniques are of particular interest because they do not require additional training processes and can directly improve model performance during inference. Test-time augmentation techniques generate multiple variants by transforming input test samples and aggregate the prediction results of all variants to reduce prediction uncertainty and improve model robustness. The core advantage is that it does not require additional training or complex hyperparameter tuning and can adapt to various existing models. Currently, there are conventional text test-time augmentation methods based on large language model generation, such as GPT3Mix or conventional large language model rewriting.

[0004] Such methods attempt to use the generation capabilities of large language models to expand data. For example, the GPT3Mix method generates new samples by mixing the text and labels of multiple real samples; or directly uses a large language model to perform synonym replacement / repetition on the input sentence to generate multiple semantically similar variants.

[0005] However, the existing test-time enhancement method based on large language models has obvious defects in the secret point identification scene, including that when mixed multiple samples are used to generate enhanced data, the secret key words may be diluted, making the secret level attribute of the sample ambiguous or even changed, introducing a large amount of noise, and thus reducing the identification accuracy; the counterfactual enhancement strategy is easy to mischange the secret core keywords, leading to sample label flipping, and the conservative synonym rewriting generates samples with high repetition, which cannot provide effective information gain; the existing methods mostly process a single sentence in isolation, which is out of the context of the entire document, and when there are pronoun references, subject omissions, etc. in the sentence, the large language model cannot accurately parse the semantics, which is easy to cause semantic drift and secret level misjudgment; the existing methods unconditionally accept all enhanced samples generated by the large language model, and do not establish an effective quality evaluation mechanism, which cannot distinguish between effective samples and invalid samples, affecting the stability and accuracy of the aggregation result.

[0006] In addition, the secret environment has strict offline and lightweight requirements for technology deployment, and the existing enhancement methods mostly rely on online large language models or heavy models, which are difficult to adapt to physically isolated secure environments, and also have the problem of high computing power cost. Therefore, how to construct a test-time enhancement scheme that takes into account semantic fidelity, sample diversity, and deployment adaptability in the few-sample and offline secret environment has become a key problem for the landing of secret point identification technology. SUMMARY

[0007] Therefore, it is necessary to provide a secret point identification method and device combining entropy guidance and large language model test enhancement to solve the problems of poor generalization performance of secret point identification in the few-sample scene, uncontrollable quality of enhanced samples, loss of context semantics, and insufficient adaptability to secret environments.

[0008] A secret point identification method combining entropy guidance and large language model test enhancement, the method comprising: Obtain a small amount of secret text data labeled with a secret level, balance the training data using oversampling technology, and fine-tune the balanced training data using a lightweight pre-trained model to obtain a classification model; Input the entire document to be identified into a locally deployed quantized large language model to generate a document summary summarizing the core content of the document, and split the entire document to be identified into an independent sentence sequence; Input the document summary and each original sentence in the sentence sequence into the quantized large language model as a prompt word, and generate a preset number of enhanced samples with the same semantics as the original sentence but different description styles according to a preset constraint; Calculate the entropy change of each original sentence before and after the corresponding enhanced sample is rewritten, judge whether the entropy change is greater than zero, if greater than zero, accept the set of enhanced samples, if not greater than zero, trigger the retry mechanism, require the quantized large language model to regenerate the enhanced sample, if the retry is preset number of times and the entropy change is still not greater than zero, give up the enhancement, keep the original sentence; The filtered enhanced sample and the corresponding original sentence are input into the classification model to obtain the classified level prediction result of each sample, and a preset aggregation strategy is used to aggregate and process multiple classified level prediction results to output the final classified level of each original sentence.

[0009] A classified level recognition device combining entropy guidance and large language model test enhancement, the device comprises: A basic model training module is configured to obtain a small amount of classified level labeled classified text data, balance the training data by using an oversampling technique, fine-tune the balanced training data by using a lightweight pre-training model, and obtain a classification model; A document preprocessing module is configured to input a whole document to be recognized in classified level into a locally deployed quantized large language model to generate a document summary summarizing the core content of the document, and split the whole document to be recognized in classified level into an independent sentence sequence; An enhanced sample generation module is configured to input the document summary and each original sentence in the sentence sequence as a prompt word into the quantized large language model, and generate a preset number of enhanced samples with different description styles but the same semantics as the original sentence according to a preset constraint; An entropy-guided enhanced sample screening module is configured to calculate the entropy change of each original sentence before and after the corresponding enhanced sample is rewritten, judge whether the entropy change is greater than zero, if greater than zero, accept the set of enhanced samples, if not greater than zero, trigger the retry mechanism, require the quantized large language model to regenerate the enhanced sample, if the retry is preset number of times and the entropy change is still not greater than zero, give up the enhancement, keep the original sentence; A classified level prediction and result aggregation module is configured to input the filtered enhanced sample and the corresponding original sentence into the classification model to obtain the classified level prediction result of each sample, and use a preset aggregation strategy to aggregate and process multiple classified level prediction results to output the final classified level of each original sentence.

[0010] A computer device comprises a memory and a processor, the memory stores a computer program, and the processor implements the following steps when executing the computer program: Obtain a small amount of classified level labeled classified text data, balance the training data by using an oversampling technique, fine-tune the balanced training data by using a lightweight pre-training model, and obtain a classification model; The entire document whose security level needs to be identified is input into the locally deployed quantized large language model to generate a document summary that summarizes the core content of the document, while the entire document whose security level needs to be identified is split into independent sentence sequences. The document summary and each original sentence in the sentence sequence are used as prompt words to input into the quantized large language model, and an augmented sample with the same semantics as the original sentence but different descriptive style is generated according to the preset constraints. Calculate the entropy change of each original sentence and its corresponding augmented sample before and after rewriting. Determine if the entropy change is greater than zero. If it is greater than zero, accept the augmented sample. If it is not greater than zero, trigger the retry mechanism and require the quantized large language model to regenerate the augmented sample. If the entropy change is still not greater than zero after a preset number of retries, abandon the augmentation and retain the original sentence. The enhanced samples that have passed the screening are input into the classification model along with the corresponding original sentences to obtain the density level prediction results of each sample; the multiple density level prediction results are aggregated using a preset aggregation strategy to output the final density level of each original sentence.

[0011] The aforementioned method and apparatus for dense point recognition, combining entropy-guided and large language model testing enhancements, utilizes oversampling technology to balance training data and fine-tunes a lightweight pre-trained model. This allows for the construction of a model with basic recognition capabilities even with limited labeled confidential data, significantly mitigating overfitting. Entropy-guided testing-time enhancements filter out highly diverse and effective samples, and aggregation strategies integrate multiple prediction results, further reducing the uncertainty of single-sample predictions. Regarding sample quality control, the generation stage employs triple constraints on format, length, and rewriting strategies to prevent semantic distortion and redundancy from the outset. The screening stage uses entropy change as a hard indicator, coupled with a three-retry fallback mechanism, to reject invalid and low-quality samples, ensuring that the enhanced samples possess substantial information gain. Simultaneously, a document summarization-assisted rewriting mechanism allows the model to perceive the global context, accurately parsing semantic information such as pronoun reference and subject omission, completely resolving semantic drift caused by detachment from context and significantly reducing the risk of misclassification of confidentiality levels. A local deployment mode is adopted, combining a large language model with a lightweight classifier, satisfying physical isolation requirements while reducing computational consumption, without relying on external online models or large-scale external data. The end-to-end closed-loop design requires no complex hyperparameter tuning and can be directly embedded into existing security management systems, adapting to the hardware conditions of different classified environments, making it highly feasible for implementation. This application possesses excellent flexibility and scalability; the aggregation strategy can be flexibly switched between voting and probability-means methods, and reasonable substitutions are supported for the basic model, large language model, and context acquisition methods. It can be adjusted and optimized according to actual scenario requirements, effectively solving four core problems: poor generalization with few samples, uncontrollable sample quality, loss of contextual semantics, and insufficient adaptability to classified environments. It provides reliable technical support for precise security management, combining practicality and innovation. Attached Figure Description

[0012] Figure 1 This is an application scenario diagram of a dense point recognition method that combines entropy guidance and large language model testing enhancement in one embodiment; Figure 2 This is a structural block diagram of a dense point recognition device that combines entropy guidance and large language model testing enhancement in one embodiment; Figure 3 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0014] In one embodiment, such as Figure 1 As shown, a dense point recognition method combining entropy guidance and large language model testing enhancement is provided, including the following steps: Step 102: Obtain a small amount of classified text data labeled with security level, use oversampling technology to balance the training data, and use a lightweight pre-trained model to fine-tune the balanced training data to obtain a classification model.

[0015] Limited-classification classified text data refers to a portion of classified text data that is restricted by confidentiality regulations and cannot be obtained on a large scale; only a portion of such classified text data, such as confidential, secret, and public, can be collected. Oversampling is a technique used to address the class imbalance problem commonly found in classified text data. By increasing the number of minority class samples, the distribution of samples across classes in the training data becomes more balanced, preventing the model from being biased towards the majority class during training and improving the model's ability to recognize various classification levels of text. The lightweight pre-trained model chosen is ernie-3.0-tiny-medium-v2-zh, which is lightweight, adaptable to deployment requirements in classified environments, and possesses good Chinese text classification capabilities. Through fine-tuning on the balanced training data, the model learns the classification-related features of classified text, thus acquiring basic classification point recognition capabilities, resulting in a classification model f that can be used for final classification determination. This solves the problem of insufficient model training due to data scarcity in classified scenarios. By oversampling to balance the data and fine-tuning the lightweight model, a model with basic recognition capabilities is constructed under limited sample conditions.

[0016] Step 104: Input the entire document whose security level needs to be identified into the locally deployed quantized large language model to generate a document summary that summarizes the core content of the document, and at the same time, split the entire document whose security level needs to be identified into independent sentence sequences.

[0017] The entire document whose security level needs to be identified refers to a classified document that requires security point identification to determine the security level of each sentence. The locally deployed quantized large language model uses a distilled quantization version of the DeepseekR1 model. Local deployment meets the physical isolation requirements of classified environments, preventing data leakage, while quantization reduces the model's computational consumption and adapts to the hardware conditions of classified environments. Generating a document summary involves using this quantized large language model to analyze the entire document and extract summary information that encapsulates the core content, providing global contextual support for subsequent sentence rewriting. Splitting the entire document into independent sentence sequences involves dividing the entire document into the form S={s1,s2,...,sn} according to the rules for dividing natural sentences, allowing each sentence to be treated as an independent recognition unit for subsequent processing. By generating a document summary and splitting the sentence sequence, a contextual basis is provided to address the semantic loss problem caused by isolated sentence processing, and preparation is made for subsequent sentence-by-sentence enhancement and recognition. Simultaneously, the locally deployed quantized model ensures the security and adaptability of the processing.

[0018] Step 106: Input the document summary and each original sentence in the sentence sequence as prompt words into the quantization large language model, and generate a preset number of augmented samples with the same semantics as the original sentences but different descriptive styles according to preset constraints.

[0019] The specific structure of the prompt words includes an instruction area, an example area, and an input area. The input area explicitly contains two core components: a document summary and the original sentence, enabling the quantized large language model to simultaneously acquire local sentence information and global document context. Preset constraints include format constraints, length constraints, and rewriting strategy constraints. The format constraint requires the output format of the augmented samples to be a JSON array without any extra content, facilitating automated parsing. The length constraint requires the length of the rewritten augmented samples to be similar to the length of the original sentence, avoiding semantic distortion due to significant length changes. The rewriting strategy constraint prompts the quantized large language model to use one or more methods from lexical rewriting, syntactic reconstruction, and style conversion. Examples include synonym replacement at the lexical level, sentence structure adjustment at the syntactic level, and formal vs. colloquial style conversion. The preset number of samples is 4. This number ensures the diversity of the augmented samples, providing sufficient feature support for subsequent aggregation prediction, while avoiding excessive computational consumption and decreased processing efficiency due to an excessive number of samples. By incorporating document summaries into prompt words, the semantic ambiguity caused by isolated sentence processing in existing technologies is resolved. At the same time, through explicit pre-defined constraints, the enhanced samples are ensured to have stylistic diversity while maintaining semantic fidelity, providing a high-quality sample foundation for improving the robustness of model recognition.

[0020] Step 108: Calculate the entropy change of each original sentence and its corresponding augmented sample before and after rewriting. Determine if the entropy change is greater than zero. If it is, accept the augmented sample. If it is not, trigger the retry mechanism, requiring the quantized large language model to regenerate the augmented sample. If the entropy change is still not greater than zero after a preset number of retries, abandon the augmentation and retain the original sentence. Entropy change is a quantitative indicator calculated based on the statistical distribution of lexical units, used to assess the diversity difference between augmented samples and the original sentence. The preset number of retries is 3. This setting ensures that as many augmented samples as possible meet the requirements while avoiding process blockage caused by infinite retries, thus improving processing efficiency. This step uses an entropy increase judgment mechanism to quantitatively screen the quality of augmented samples, rejecting samples with insufficient diversity. This ensures that samples entering subsequent classification stages can provide effective information gain, solving the noise introduction problem caused by the unconditional acceptance of augmented samples in existing technologies. Furthermore, the retry mechanism and fallback strategy further guarantee the stability and reliability of the process.

[0021] Step 110: Input the enhanced samples that have passed the screening and the corresponding original sentences into the classification model to obtain the density level prediction results of each sample; use a preset aggregation strategy to aggregate the multiple density level prediction results and output the final density level of each original sentence.

[0022] The enhanced samples that pass the screening refer to the set of valid enhanced samples whose entropy change is greater than zero after entropy increase judgment. s' i,1 ,..., s' i,m}, m Four samples are pre-set and input into the classification model along with the original sentence. This approach fully utilizes the feature information from multiple samples, reducing the uncertainty of predictions based on a single sample. Two preset aggregation strategies are available: voting and probability mean. Voting calculates the frequency of each security level label across all predictions, selecting the label with the highest frequency as the final security level. Probability mean calculates the average predicted probability of all samples across all security levels, selecting the security level with the highest average. Both aggregation strategies effectively integrate the prediction information from multiple samples, improving the accuracy and stability of the final security level determination. This step, through multi-sample input and aggregation strategy processing, further reduces the risk of security level misclassification, significantly improving the model's generalization performance in scenarios with few samples.

[0023] The aforementioned method and apparatus for dense point recognition, combining entropy-guided and large language model testing enhancements, utilizes oversampling technology to balance training data and fine-tunes a lightweight pre-trained model. This allows for the construction of a model with basic recognition capabilities even with limited labeled confidential data, significantly mitigating overfitting. Entropy-guided testing-time enhancements filter out highly diverse and effective samples, and aggregation strategies integrate multiple prediction results, further reducing the uncertainty of single-sample predictions. Regarding sample quality control, the generation stage employs triple constraints on format, length, and rewriting strategies to prevent semantic distortion and redundancy from the outset. The screening stage uses entropy change as a hard indicator, coupled with a three-retry fallback mechanism, to reject invalid and low-quality samples, ensuring that the enhanced samples possess substantial information gain. Simultaneously, a document summarization-assisted rewriting mechanism allows the model to perceive the global context, accurately parsing semantic information such as pronoun reference and subject omission, completely resolving semantic drift caused by detachment from context and significantly reducing the risk of misclassification of confidentiality levels. A local deployment mode is adopted, combining a large language model with a lightweight classifier, satisfying physical isolation requirements while reducing computational consumption, without relying on external online models or large-scale external data. The end-to-end closed-loop design requires no complex hyperparameter tuning and can be directly embedded into existing security management systems, adapting to the hardware conditions of different classified environments, making it highly feasible for implementation. This application possesses excellent flexibility and scalability; the aggregation strategy can be flexibly switched between voting and probability-means methods, and reasonable substitutions are supported for the basic model, large language model, and context acquisition methods. It can be adjusted and optimized according to actual scenario requirements, effectively solving four core problems: poor generalization with few samples, uncontrollable sample quality, loss of contextual semantics, and insufficient adaptability to classified environments. It provides reliable technical support for precise security management, combining practicality and innovation.

[0024] In one embodiment, the entropy change of each original sentence and its corresponding augmented sample before and after rewriting is calculated, including: Calculate the entropy change of each original sentence and its corresponding augmented sample before and after rewriting.

[0025] in, N s The total number of lexical units in the original sentence. C i For the first sentence in the original sentence i The number of times each word element appears. N p The total number of lexical units for all rewritten sentences. C j ' To rewrite the sentence in the first... j The number of times each word element appears. k Indicates the number of different word types in the original sentence. mThis indicates the number of different word types in all rewritten enhanced samples.

[0026] Specifically, a lexical unit refers to the basic linguistic unit obtained after segmenting a sentence. By statistically analyzing the frequency and types of lexical units in the original and rewritten sentences, and substituting these values ​​into the formula above, the entropy values ​​before and after rewriting can be calculated, thus yielding the entropy change. The entropy value reflects the degree of disorder in the distribution of lexical units in the text, while the entropy change... >0 indicates that the rewritten enhanced sample word distribution is more diverse, which can provide more diverse feature information for the model and has effective information gain; A value ≤0 indicates insufficient diversity in the augmented samples, failing to provide new and effective features for model recognition, and potentially introducing redundant information. Therefore, the samples in this group should be rejected, triggering a retry mechanism. This entropy increase calculation method provides a clear and quantifiable standard for evaluating the quality of augmented samples, solving the problem of existing technologies lacking effective sample quality evaluation indicators. It ensures that the augmented samples entering subsequent classification stages possess substantial diversity, providing a reliable guarantee for improving model recognition performance.

[0027] In one embodiment, the preset constraints include format constraints, length constraints, and rewriting strategy constraints; the format constraints require that the output format of the augmented sample be a JSON array and that it does not contain any additional content; the length constraints require that the length of the rewritten augmented sample be similar to the length of the original sentence; the rewriting strategy constraints suggest that the quantized large language model use one or more of the following methods for rewriting: lexical rewriting, syntactic reconstruction, and style transfer.

[0028] Specifically, the format constraints require the output to be a JSON array. For example, if the original sentence is "She has a beautiful face," the output format after generating four augmented samples will be ["She has a beautiful face", "She has a beautiful face", "A beautiful face is on her", "Her face is very pretty"]. This format must not contain any explanatory or descriptive text, ensuring the program can directly parse and extract the augmented samples, thus improving processing efficiency. The length constraints require the length of the rewritten sentence to deviate from the original sentence length within a reasonable range, avoiding semantic changes due to excessive expansion or reduction. For example, if the original sentence is "She works very hard," it cannot be rewritten as "Throughout her long career, she has always maintained a diligent and hardworking attitude, never slacking off, and wholeheartedly devoting herself to various work tasks" (too long) or "She works diligently" (too short), ensuring the integrity and consistency of semantic expression. The rewriting strategy constraints provide specific rewriting directions. Lexical rewriting can use synonym replacement, such as replacing "pretty" with "beautiful" or "graceful." Syntactic reconstruction can adjust sentence word order, such as rewriting "She has a pretty face" as "A pretty face grows on her body." Style conversion can switch between formal and colloquial expressions, such as rewriting "He was too tired, so he stopped working" as "Because he was too tired, he stopped working." These pre-defined constraints regulate the augmented sample generation process from three dimensions: format, length, and content. This ensures both the parsability and semantic fidelity of the samples, while also providing a clear path to sample diversity, effectively avoiding the problems of semantic distortion or insufficient diversity in augmented samples in existing technologies.

[0029] In one embodiment, the preset number of copies is 4, and the preset number of retries is 3.

[0030] In one embodiment, the preset aggregation strategy is a voting method, which includes counting the number of times each security level label appears in the prediction results and taking the security level label with the most occurrences as the final security level of the corresponding original sentence.

[0031] Specifically, assuming an original sentence and its four augmented samples are input into a classification model, the predicted security level results are "Confidential", "Confidential", "Secret", "Confidential", and "Confidential". The frequency of each security level label is counted: "Confidential" appears 4 times, and "Secret" appears once. The most frequent "Confidential" is then taken as the final security level of the original sentence. The core idea of ​​the voting method is to use the prediction results of the majority of samples to offset the prediction errors of individual samples. Since the augmented samples are generated under the premise of semantic fidelity, their prediction results have high reliability. By using a majority vote, the randomness of single-sample predictions can be reduced, improving the accuracy and stability of the final security level determination. This aggregation strategy is simple, intuitive, and easy to implement, requiring no complex probability calculations. While ensuring recognition accuracy, it reduces computational complexity, adapts to the lightweight deployment requirements of classified environments, and effectively solves the problem of insufficient reliability of single prediction results in existing technologies.

[0032] In one embodiment, the preset aggregation strategy is replaced by the probability mean method, which includes calculating the mean predicted probability of all enhanced samples and the original sentence at various security levels, and taking the security level with the largest mean as the final security level of the corresponding original sentence.

[0033] Specifically, the probability mean method fully utilizes the probability information output by the classification model, reflecting the confidence level of each sample's prediction results better than the voting method. By calculating the mean, it smooths out extreme prediction results for individual samples, further improving the reliability and stability of the final judgment. This method provides a flexible alternative to aggregation strategies, allowing selection based on the needs of the actual confidential scenario and the characteristics of the model output, thus enhancing the adaptability and flexibility of the technical solution.

[0034] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0035] In one embodiment, such as Figure 2As shown, a dense point recognition device combining entropy guidance and large language model testing enhancement is provided, including: a basic model training module 202, a document preprocessing module 204, an enhanced sample generation module 206, an entropy-guided enhanced sample screening module 208, and a density level prediction and result aggregation module 210, wherein: The basic model training module 202 is used to acquire a small amount of classified text data labeled with a security level, use oversampling technology to balance the training data, and use a lightweight pre-trained model to fine-tune the balanced training data to obtain a classification model. The document preprocessing module 204 is used to input the entire document whose security level needs to be identified into the locally deployed quantized large language model, generate a document summary that summarizes the core content of the document, and at the same time split the entire document whose security level needs to be identified into independent sentence sequences. The augmented sample generation module 206 is used to input the document summary and each original sentence in the sentence sequence as prompt words into the quantization large language model, and generate a preset number of augmented samples with the same semantics as the original sentences but different descriptive styles according to preset constraints. The entropy-guided enhancement sample selection module 208 is used to calculate the entropy change of each original sentence and its corresponding enhancement sample before and after rewriting, and to determine whether the entropy change is greater than zero. If it is greater than zero, the enhancement sample is accepted; if it is not greater than zero, a retry mechanism is triggered, requiring the quantized large language model to regenerate the enhancement sample. If the entropy change is still not greater than zero after a preset number of retries, the enhancement is abandoned and the original sentence is retained. The security level prediction and result aggregation module 210 is used to input the filtered enhanced samples and the corresponding original sentences into the classification model to obtain the security level prediction results of each sample; and to aggregate multiple security level prediction results using a preset aggregation strategy to output the final security level of each original sentence.

[0036] Specific limitations regarding the dense point recognition device combining entropy guidance and large language model testing enhancement can be found in the above-described limitations of the dense point recognition method combining entropy guidance and large language model testing enhancement, and will not be repeated here. Each module in the aforementioned dense point recognition device combining entropy guidance and large language model testing enhancement can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0037] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 3As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When executed by the processor, the computer program implements a dense point recognition method that combines entropy guidance and large language model testing enhancement. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0038] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0039] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0040] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0041] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for point cloud recognition combining entropy guidance and large language model test enhancement, characterized in that, The method comprises: Obtaining a small amount of classified text data with marked classification levels, balancing the training data by using an oversampling technique, fine-tuning the balanced training data by using a lightweight pre-trained model to obtain a classification model; Inputting the whole document to be identified into a locally deployed quantized large language model to generate a document summary summarizing the core content of the document, and meanwhile, splitting the whole document to be identified into an independent sentence sequence; Inputting the document summary and each original sentence in the sentence sequence into the quantized large language model as prompt words, and generating a preset number of enhanced samples with the same semantics as the original sentence but different description styles according to preset constraints; Calculating the entropy change amount of each original sentence before and after the corresponding enhanced sample is rewritten, judging whether the entropy change amount is greater than zero, if yes, accepting the enhanced sample, if no, triggering a retry mechanism to require the quantized large language model to re-generate the enhanced sample, and if the retry number reaches a preset number and the entropy change amount is still not greater than zero, giving up the enhancement and keeping the original sentence; Inputting the screened enhanced sample and the corresponding original sentence into the classification model to obtain the classification level prediction result of each sample, and adopting a preset aggregation strategy to aggregate and process a plurality of classification level prediction results to output the final classification level of each original sentence.

2. The method of claim 1, wherein, The method comprises: The preset constraints comprise format constraints, length constraints and rewriting strategy constraints; the format constraints require that the output format of the enhanced sample is a JSON array and does not contain additional content; wherein, N s is the total number of wordpieces in the original sentence, C i is the number of occurrences of the wordpiece in the original sentence, i N p is the total number of wordpieces in all the paraphrased sentences, C j ' is the number of occurrences of the wordpiece in the paraphrased sentence, j represents the number of different wordpiece types in the original sentence, represents the number of different wordpiece types in all the paraphrased enhanced samples.​​ 3. The method of claim 1, wherein, The length constraints require that the length of the rewritten enhanced sample is similar to that of the original sentence; and the rewriting strategy constraints prompt the quantized large language model to adopt one or more methods of vocabulary rewriting, syntax reconstruction and style conversion for rewriting. The preset number is 4, and the preset retry number is 3.

4. The method of claim 1, wherein, The preset aggregation strategy is a voting method, which comprises counting the number of times each classification level label appears in the prediction result, and taking the classification level label with the maximum number of appearances as the final classification level of the corresponding original sentence.

5. The method of claim 1, wherein, The preset aggregation strategy is replaced by a probability mean method, which comprises calculating the mean value of the prediction probability of all enhanced samples and original sentences on each classification level, and taking the classification level with the maximum mean value as the final classification level of the corresponding original sentence.

6. The method of claim 1, wherein, The device comprises:

7. A device for identifying a point of interest, combining entropy guidance and large language model test enhancement, characterized in that, A basic model training module configured to obtain a small amount of classified text data with marked classification levels, balance the training data by using an oversampling technique, fine-tune the balanced training data by using a lightweight pre-trained model to obtain a classification model; A document preprocessing module configured to input the whole document to be identified into a locally deployed quantized large language model to generate a document summary summarizing the core content of the document, and meanwhile, split the whole document to be identified into an independent sentence sequence; An enhanced sample generation module configured to input the document summary and each original sentence in the sentence sequence into the quantized large language model as prompt words, and generate a preset number of enhanced samples with the same semantics as the original sentence but different description styles according to preset constraints; An enhanced sample generation module configured to input the document summary and each original sentence in the sentence sequence into the quantized large language model as prompt words, and generate a preset number of enhanced samples with the same semantics as the original sentence but different description styles according to preset constraints; The entropy-guided enhanced sample screening module is configured to calculate an entropy change amount of each original sentence before and after the original sentence is rewritten by the corresponding enhanced sample, determine whether the entropy change amount is greater than zero, accept the set of enhanced samples if the entropy change amount is greater than zero, and trigger a retry mechanism if the entropy change amount is not greater than zero, requiring the quantitative large language model to re-generate the enhanced sample. If the entropy change amount is still not greater than zero after retrying for a preset number of times, the enhancement is abandoned, and the original sentence is retained. The secret level prediction and result aggregation module is configured to input the screened enhanced sample and the corresponding original sentence into the classification model to obtain a secret level prediction result of each sample, and perform aggregation processing on a plurality of secret level prediction results by using a preset aggregation strategy to output a final secret level of each original sentence. 8.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-7. The processor implements the steps of the method in any one of claims 1 to 6 when executing the computer program.

Images