Aquaculture disease prevention and control text data enhancement method, device and storage medium based on small model guidance

Abstract

The present invention provides a method, device and storage medium for enhancing text data for aquaculture disease prevention and control based on small model guidance, which belongs to the field of text data enhancement. The method is based on the strategy of high-quality corpus + pre-trained large model + prompt, guiding the model to generate data that meets the requirements; introducing a small model supervision large model method, combining the small model with reinforcement learning technology, using the small model to capture the semantic information in the text, and through the reinforcement learning mechanism, continuously and dynamically adjusting the generation strategy of the large model according to the feedback of the small model, thereby improving the relevance of the generated data and enhancing the diversity of the data; in addition, a noise removal integration module is designed, using the small model to be responsible for evaluating and screening the data generated by the large model, identifying and excluding data that does not meet the requirements or contains noise, thereby improving the quality of the overall data set. The text generated by the present invention not only maintains a high degree of consistency with the original data in terms of semantics, but also significantly increases the richness of the data.

Description

Technical Field

[0001] The present invention belongs to the technical field of text data enhancement, and relates to a text data enhancement method for aquaculture disease prevention and control based on small model guidance, an electronic device and a storage medium. Background Art

[0002] Aquaculture is crucial to global food security and economic progress, but disease outbreaks often lead to mass mortality events, resulting in severe economic losses for aquaculture farmers. For example, certain infectious diseases can spread rapidly, causing mass mortalities in fish or shellfish, leading to growth retardation, reduced feed conversion rates, and increased production costs. Therefore, introducing advanced technologies and innovative solutions to aquaculture to achieve rapid and accurate disease diagnosis is crucial for improving efficiency and economic benefits. However, practical implementation presents numerous challenges. First, the complex and dynamic aquaculture environment, with significant differences across regions, seasons, and species, creates a lack of standardized recording methods and assessment systems, making data unification difficult. Second, many small farms lack comprehensive recording systems and technical support, resulting in limited available data. These challenges not only impact the accuracy of data analysis but also hinder the development and application of data-driven decision support systems. In summary, addressing the scarcity and consistency of data in aquaculture disease prevention and control is a core task in promoting technological advancement in this field.

[0003] Data augmentation methods can alleviate data scarcity in the field of aquatic disease control. Traditional data augmentation methods, such as random word insertion or deletion, synonym-based semantic substitution, and back-translation, often introduce noise, compromise semantic integrity, and lack richness. With the development of large language models (LLMs), new data augmentation methods have emerged. For example, GPT3Mix increases data diversity by mixing real samples, but its randomness can compromise sample authenticity. AugGPT utilizes large models to paraphrase sentences to ensure conceptual consistency, but this results in lower richness. Current data augmentation methods struggle to simultaneously ensure both accuracy and diversity in generated data, especially in the field of aquatic disease control, where accuracy is paramount. Large language models can produce "hallucinations" when processing specialized content, providing inaccurate information. This is particularly problematic in fields requiring high precision. For example, the term "infectious" has different meanings across different fields, highlighting the importance of understanding specific terminology. Therefore, in the field of aquatic disease control, in addition to increasing data diversity, special attention must be paid to improving data quality and applicability to ensure the validity and reliability of generated content. This is not only the key to meeting current challenges, but also the basis for promoting the precise application of large models in the prevention and control of aquatic diseases.

[0004] To address the lack of domain-specific knowledge in large language models, researchers have explored fine-tuning strategies such as low-rank adaptation (LoRA), P-tuning v2, and freezing methods, aiming to effectively incorporate domain expertise into large models and improve the accuracy of generated content. However, improving accuracy alone is not enough to ensure that model outputs meet user needs. Therefore, human-feedback reinforcement learning (RLHF) methods, such as proximal policy optimization (PPO), direct preference optimization (DPO), and k-to-transform (KTO), have been introduced to integrate user feedback into the training process to better align the model with user expectations. However, as the complexity of model tasks increases, evaluation and supervision become more difficult, especially in high-risk industries such as healthcare, where traditional manual feedback mechanisms struggle to meet the accuracy and efficiency requirements. In summary, although large models show great potential for data augmentation, their application in the field of aquatic diseases still faces challenges such as designing high-quality data augmentation strategies and ensuring the controllability of model-generated data. Summary of the Invention

[0005] In order to solve the problem that the amount of data in the field of aquatic diseases is small and the existing data augmentation methods are difficult to simultaneously ensure the accuracy and diversity of the generated data, the present invention proposes a text data augmentation method for aquaculture disease prevention and control based on small model guidance. This method is based on the strategy of "high-quality corpus + pre-trained large model + efficient prompt", and improves the data augmentation effect by integrating high-quality corpus and optimizing prompt technology; then, in order to address the problem that the quality of augmented data is not high due to the poor adaptability of large language models in the field of aquatic biological disease prevention and control, a supervised learning model is introduced to guide the large model's understanding of domain knowledge and the optimization of data augmentation tasks; in addition, a denoising module is designed to eliminate noise by analyzing the consistency between the augmented data and the data augmentation target, thereby further improving the overall quality of the augmented data; the present invention achieves data with high similarity and high richness to the original professional content.

[0006] The technical solutions of the present invention are as follows:

[0007] A small model-guided text data enhancement method for aquaculture disease prevention and control, comprising:

[0008] S1. Construct the original sample set of aquatic biological diseases and perform fine-tuning on the corpus annotation to generate the aquatic biological disease annotation dataset;

[0009] S2. Split each piece of data in the original sample set into a group of data segments using sentence-level segmentation;

[0010] S3, build a Prompt framework that integrates prompt engineering and rule checking strategies;

[0011] S4: Based on the aquatic biological disease annotated dataset constructed in S1, supervised fine-tuning is performed on the pre-trained large model to obtain the SFT domain large model; and the Prompt framework constructed in S3 is used to guide the SFT domain large model to generate enhanced data;

[0012] S5. Use the small model as a supervision model, input the slice data obtained in S2 and the enhanced data obtained in S4 into the supervision model, and use the small model supervision large model (SMS-KTO) training method to train the SFT domain large model to obtain a large model for data augmentation in the field of aquatic diseases;

[0013] S6. Use the data augmentation model in the field of aquatic diseases obtained in S5 to generate enhanced data, and perform noise removal and aggregation processing on the enhanced data to obtain the final long text data.

[0014] Furthermore, the specific process of S1 is as follows:

[0015] S1.1. Search using the names of aquatic biological diseases as keywords and collect information on several attribute types corresponding to each disease name to form the original sample set D. The attribute types include disease type, pathogen / cause, target of harm, main symptoms, prevalence, treatment, control measures, and prevention methods.

[0016] S1.2. Use the attribute-value pair data annotation method in structured annotation to annotate the aquatic biological disease data in the original sample set D; each aquatic biological disease data s in the original sample set D i By the core entity E t And its eight entity attributes E a (a≤8), where the core entity E t That is, the name of the disease, entity attribute E a That is, the attribute type, each entity attribute includes m attribute descriptions, so each aquatic biological disease data is represented by s i ={E t ,E a (x1,x2,...x m )}, annotated as {E t ,(E1(x1),y 1,1 ),(E1(x2),y 1,2 ),...,(E a (x m ),y a,m )}, where y a,m Represents labeled data. All aquatic biological disease data are labeled according to this logic to form the labeled dataset Z.

[0017] Furthermore, the specific method of S2 is:

[0018] For a piece of aquatic biological disease data in the original sample set D, the core entity E is first extracted. t (i.e. disease name) and related entity attributes E a ; Then the original data is segmented at the sentence level according to entity attributes to form a series of new data records S i ; Each S i By a specific core entity E t The corresponding entity attribute E a This process is expressed as:

[0019] D s =Split Sentence Level(D)={S1,S2,S3,S4,...} (1)

[0020] Among them, Split Sentence Level represents the operation of splitting the original data according to the sentence level; each S i =E t +E a Represents the i-th sentence-level data segment split from the original sample set D.

[0021] Furthermore, in the S3, the construction process of the Prompt framework includes:

[0022] S3.1. To ensure that the content generated by the large language model meets both the intended purpose and the specified format requirements, a "purpose + requirements + format" framework is adopted. Let P be the final generated prompt word, M be the goal (i.e., purpose) of the generated content, R be the requirements for the generated content, and F be the format of the output content. The prompt word design process can be expressed as follows:

[0023] P=Prompt Design(M,R,F) (2)

[0024] The Prompt Design function represents the prompt word design process. It accepts three parameters: target M, requirement R, and format F, and outputs an optimized prompt word P.

[0025] S3.2. After generating the content, an automated mechanism is used to check each generated data item by item to see if it meets the predetermined requirements. If it does not meet the initial set requirements, it will be regenerated until the threshold conditions are met. The entire process can be expressed as follows:

[0026] C final =Generate And Validate(N,T,V) (3)

[0027] The workflow of the Generate And Validate function is as follows:

[0028]

[0029] Among them, C final represents the final generated content; N is the number of contents to be generated; T is the threshold condition, i.e. the maximum number of attempts; C t Indicates the content generated for the tth time; V(C t ) is a verification function used to check C t Whether the predetermined standards are met.

[0030] Furthermore, the specific process of S4 is as follows:

[0031] The aquatic biological disease annotation dataset Z is used to perform LORA supervised fine-tuning on the parameters θ in the pre-trained large model. After fine-tuning, the SFT domain large model is obtained as follows:

[0032] h student =W0x+ΔWx=W0x+BAx (5)

[0033] Using the Prompt framework to guide the large model in the SFT field to generate enhanced data L(s i ):

[0034]

[0035] Wherein, P represents the prompt word.

[0036] Furthermore, the pre-trained large model includes ChatGLM model, LLaMA model, Qwen model, etc.

[0037] Furthermore, the specific process of S5 is as follows:

[0038] S5.1. Construct a two-level supervision system in the supervision model, where one level focuses on the similarity evaluation between sentences and the other level focuses on the richness analysis of information. i ) and the original slice data S obtained by S2 i Input into the supervised model to generate multi-objective reward scores, namely similarity scores and richness scores;

[0039] The original slice data S i Embedding v and augmented data L(s i ) Sentence embedding u i Perform cosine distance calculation to obtain the similarity score Q s :

[0040]

[0041] Richness score Q r The calculation formula is as follows:

[0042]

[0043] Among them, J[L(s i ),S i ] represents the Jaccard distance, which is used to describe L(s i ) and S i The difference between them, that is, richness;

[0044] S5.2, corresponding to the two reward objectives of similarity score and richness score, two gating layers are introduced into the supervised model to comprehensively consider the balance between the two dimensions and their relationship with the enhanced data; specifically, input Q s and its threshold condition Z, and Q r and its threshold condition H, the output G(Q s ,Q r ) is expressed as:

[0045] G(Q s ,Q r )=|(Q s ∈Z)|(Q r ∈H) (10)

[0046] In this way, it is possible to accurately determine which examples have a reward target score above or below a preset threshold;

[0047] S5.3, binary label G generated by formula (10), prompt word P (s i ) and enhanced data L(s i ) is used to form a data set, and the parameters of the large model in the SFT field are trained based on the KTO reinforcement learning method using this data set to obtain a large model for data augmentation in the field of aquatic diseases. The loss function of KTO training is defined as follows:

[0048]

[0049] in:

[0050]

[0051] Among them, z ref is the KL divergence term, r KTO is the reward function form derived in DPO; π θ is the model being aligned, π ref is the reference model, σ is the logarithmic function, λ D and λ Uare the hyperparameters of the desired output and the undesirable output, respectively;

[0052] The training parameters of the SFT domain large model are as follows:

[0053]

[0054] Among them, γ is a hyperparameter, This is the model being aligned after LORA fine-tuning using D. is a trainable parameter.

[0055] Furthermore, the specific process of S6 is as follows:

[0056] S6.1. Assume that the set of all data generated by the data augmentation model in the field of aquatic diseases is D':D'→R 2 is a mapping function that generates a two-dimensional vector (Q s (d),Q r (d)) represents the similarity score and richness score respectively, and the optimal selection is made layer by layer, and the optimal ratio is adjusted dynamically to finally obtain high-quality data D f , the preliminary screening process is expressed as:

[0057] D f ={d∈D'|Q s (d)≥τ d ∩Q r (d)≥τ r} (17)

[0058] Among them, τ d Represents the similarity screening index, τ r represents the richness screening index;

[0059] S6.2, in order to f Restore the original long text knowledge using a random extraction and aggregation method: Assume D s ={S1,S2,S3,S4,...} is the original slice data set, where each element S i is a single sentence containing a single attribute; It's D s Medium S i After data enhancement and processing (such as denoising, data enhancement, etc.), the data set is Where n∈R represents the number of data augmentations; for D' f Augmented data of the same attribute in Random extraction is performed to form long text augmentation data. The specific implementation is as follows:

[0060]

[0061] Among them, aug is the aggregation function, Random Sample(D' f ,n) indicates that from D' f Each Randomly extract the corresponding n augmented versions and merge them into a long text S aug,N ; Can generate long text augmented data D aug The amount of data C total =n i .

[0062] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor; when the processor executes the computer program, the electronic device executes the method described in any of the above embodiments.

[0063] The present invention also provides a storage medium, which includes a computer program. When the computer program is run on an electronic device, the electronic device executes the method described in any of the above embodiments.

[0064] Beneficial effects of the present invention: The present invention combines the powerful generative capabilities of large-scale pre-trained language models and the precise guidance of small-scale supervised models, and achieves a high degree of supervision of generated data by introducing prompt learning and reinforcement learning techniques, ensuring that the generated text is not only highly consistent with the original data in semantics, but also significantly increases the richness of the data. Experimental results show that compared with various existing data enhancement methods, the classification effect of the present invention on data sets in different fields is better than the baseline model. The generated samples are not only highly similar to the original data in semantics, but also have higher diversity in content. This balance enables the model to learn features from more angles, thereby improving its robustness and generalization ability. The present invention can promote faster domain adaptation and better few-sample learning of generative models in computer vision, and realize multimodal data augmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 It is an overall flow chart of the method of the present invention.

[0066] Figure 2 It is a flowchart of the small model supervising the large model.

[0067] Figure 3 is the confusion matrix constructed when training using the original data.

[0068] Figure 4 It is the confusion matrix constructed when the model is trained after data augmentation using the method of the present invention. DETAILED DESCRIPTION

[0069] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0070] like Figure 1 As shown, the embodiment of the present invention provides a method for enhancing text data of aquaculture disease prevention and control based on a small model, comprising the following steps:

[0071] S1. Construct the original sample set of aquatic biological diseases and perform fine-tuning on the corpus annotation to generate the aquatic biological disease annotated dataset. The specific process is as follows:

[0072] S1.1. Search using the names of aquatic biological diseases as keywords and collect information on several attribute types corresponding to each disease name to form the original sample set D. The attribute types include disease type, pathogen / cause, target of harm, main symptoms, prevalence, treatment, control measures, and prevention methods.

[0073] S1.2. Use the attribute-value pair data annotation method in structured annotation to annotate the aquatic biological disease data in the original sample set D; each aquatic biological disease data s in the original sample set D i By the core entity E t And its eight entity attributes E a (a≤8), where the core entity E t That is, the name of the disease, entity attribute E a That is, the attribute type, each entity attribute includes m attribute descriptions, so each aquatic biological disease data is represented by s i ={E t ,E a (x1,x2,...x m )}, annotated as {E t ,(E1(x1),y 1,1 ),(E1(x2),y 1,2 ),...,(E a (x m ),y a,m )}, where y a,m Represents labeled data. All aquatic biological disease data are labeled according to this logic to form the labeled dataset Z.

[0074] S2. Split each piece of data in the original sample set into a group of data segments using sentence-level segmentation. The specific method is as follows:

[0075] For a piece of aquatic biological disease data in the original sample set D, the core entity E is first extracted. t(i.e. disease name) and related entity attributes E a ; Then the original data is segmented at the sentence level according to entity attributes to form a series of new data records S i ; Each S i By a specific core entity E t The corresponding entity attribute E a This process is expressed as:

[0076] D s =Split Sentence Level(D)={S1,S2,S3,S4,...} (1)

[0077] Among them, Split Sentence Level represents the operation of splitting the original data according to the sentence level; each S i =E t +E a Represents the i-th sentence-level data segment split from the original sample set D.

[0078] S3. Build a Prompt framework that integrates prompt engineering and rule checking strategies. The construction process includes:

[0079] S3.1. To ensure that the content generated by the large language model meets both the intended purpose and the specified format requirements, a "purpose + requirements + format" framework is adopted. Let P be the final generated prompt word, M be the goal (i.e., purpose) of the generated content, R be the requirements for the generated content, and F be the format of the output content. The prompt word design process can be expressed as follows:

[0080] P=Prompt Design(M,R,F) (2)

[0081] The Prompt Design function represents the prompt word design process. It accepts three parameters: target M, requirement R, and format F, and outputs an optimized prompt word P.

[0082] S3.2. After generating the content, an automated mechanism is used to check each generated data item by item to see if it meets the predetermined requirements. If it does not meet the initial set requirements, it will be regenerated until the threshold conditions are met. The entire process can be expressed as follows:

[0083] C final =Generate And Validate(N,T,V) (3)

[0084] The workflow of the Generate And Validate function is as follows:

[0085]

[0086] Among them, C final represents the final generated content; N is the number of contents to be generated; T is the threshold condition, i.e. the maximum number of attempts; C t Indicates the content generated for the tth time; V(C t ) is a verification function used to check C t Whether the predetermined standards are met.

[0087] S4: Based on the aquatic biological disease annotated dataset constructed in S1, the ChatGLM model is fine-tuned in a supervised manner to obtain a large SFT domain model. The Prompt framework constructed in S3 is then used to guide the SFT domain model to generate enhanced data. The specific process is as follows:

[0088] The parameters θ in the ChatGLM model are fine-tuned by LORA supervision using the aquatic biological disease annotated dataset Z. After fine-tuning, the large model in the SFT field is as follows:

[0089] h student =W0x+ΔWx=W0x+BAx (5)

[0090] Using the Prompt framework to guide the large model in the SFT field to generate enhanced data L(s i ):

[0091]

[0092] Wherein, P represents the prompt word.

[0093] S5. Using the sentence-transformers model as a supervision model, the SMS-KTO training method is used to train the SFT domain large model to obtain a data augmentation large model for aquatic diseases. The specific process is as follows:

[0094] S5.1. Construct a two-level supervision system in the sentence-transformers model, where one level focuses on the similarity evaluation between sentences and the other level focuses on the richness analysis of information. i ) and the original slice data S obtained by S2 i Input into the supervised model to generate multi-objective reward scores, namely similarity scores and richness scores;

[0095] The original slice data S i Embedding v and augmented data L(s i ) Sentence embedding u i Perform cosine distance calculation to obtain the similarity score Q s:

[0096]

[0097] Richness score Q r The calculation formula is as follows:

[0098]

[0099] Among them, J[L(s i ),S i ] represents the Jaccard distance, which is used to describe L(s i ) and S i The difference between them, that is, richness;

[0100] S5.2, corresponding to the two reward objectives of similarity score and richness score, two gating layers are introduced into the supervised model to comprehensively consider the balance between the two dimensions and their relationship with the enhanced data; specifically, input Q s and its threshold condition Z, and Q r and its threshold condition H, the output G(Q s ,Q r ) is expressed as:

[0101] G(Q s ,Q r )=|(Q s ∈Z)|(Q r ∈H) (10)

[0102] S5.3, binary label G generated by formula (10), prompt word P (s i ) and enhanced data L(s i ) is used to form a data set, and the parameters of the large model in the SFT field are trained based on the KTO reinforcement learning method using this data set to obtain a large model for data augmentation in the field of aquatic diseases. The loss function of KTO training is defined as follows:

[0103]

[0104] in:

[0105]

[0106] Among them, z ref is the KL divergence term, r KTO is the reward function form derived in DPO; π θ is the model being aligned, π ref is the reference model, σ is the logarithmic function, λ D and λ U are the hyperparameters of the desired output and the undesirable output, respectively;

[0107] The training parameters of the SFT domain large model are as follows:

[0108]

[0109] Among them, γ is a hyperparameter, This is the model being aligned after LORA fine-tuning using D. is a trainable parameter.

[0110] S6. Generate enhanced data using the aquatic disease data augmentation model obtained in S5, and perform noise removal and aggregation processing on the enhanced data to obtain the final long text data. The specific process is as follows:

[0111] S6.1. Assume that the set of all data generated by the data augmentation model in the field of aquatic diseases is D':D'→R 2 is a mapping function that generates a two-dimensional vector (Q s (d),Q r (d)) represents the similarity score and richness score respectively, and the optimal selection is made layer by layer, and the optimal ratio is adjusted dynamically to finally obtain high-quality data D f , the preliminary screening process is expressed as:

[0112] D f ={d∈D'|Q s (d)≥τ d ∩Q r (d)≥τ r} (17)

[0113] Among them, τ d Represents the similarity screening index, τ r represents the richness screening index;

[0114] S6.2, in order to f Restore the original long text knowledge using a random extraction and aggregation method: Assume D s ={S1,S2,S3,S4,...} is the original slice data set, where each element S i is a single sentence containing a single attribute; It's D s Medium S i After data enhancement and processing (such as denoising, data enhancement, etc.), the data set is Where n∈R represents the number of data augmentations; for S' f Augmented data of the same attribute in Random extraction is performed to form long text augmentation data. The specific implementation is as follows:

[0115]

[0116] Among them, aug is the aggregation function, Random Sample(D' f ,n) indicates that from D' f Each Randomly extract the corresponding n augmented versions and merge them into a long text S aug,N ; Can generate long text augmented data D aug The amount of data C total =n i .

[0117] Experimental verification

[0118] 1. Build domain datasets

[0119] Two classification datasets are used for experiments: the aquatic fishery disease prevention and control professional field classification dataset ABDCD, and the open domain classification dataset THUCNews.

[0120] 1) The Aquatic Biological Disease Classification Dataset (ABDCD) uses self-developed aquatic biological disease classification data. In consultation with domain experts, the system was divided into seven key attributes, centered around disease name, including disease type and pathogen or cause, target, primary symptoms, prevalence, treatment, control measures, and prevention methods. Knowledge was collected through extensive searches of specialized websites, literature, and books. This specialized dataset is not only relatively small (as shown in Table 1), but also contains a large amount of industry-specific terminology, posing significant challenges for effective parsing using traditional text analysis tools. Further complicating matters, the data is highly unevenly distributed: disease type and pathogen or cause accounts for 17.4%, target accounts for 18.4%, primary symptoms account for 18%, prevalence accounts for 17.3%, treatment accounts for 11.5%, control measures account for 2.1%, and prevention methods account for 15.2%. This significant imbalance in the original data further exacerbates the difficulty of classification.

[0121] Table 1

[0122]

[0123] To construct a classification dataset: First, for each attribute, 100 samples (base samples) were randomly selected as the training set, and 100 samples (base samples) were randomly selected as the test set. Then, all remaining real samples for each attribute were added to the training set. Given the limited sample size and highly uneven distribution of classes in the dataset, a strategy was adopted to add varying proportions of generated samples (referred to as augmented samples) to the base training samples. This approach aims to augment the training set with synthetic data, allowing for a systematic evaluation of the impact of augmented samples of varying sizes on the model's attribute classification performance.

[0124] 2) THUCNews (Tsinghua News Corpus) (Sun et al., 2016) is a high-quality, large-scale Chinese news dataset compiled by the Natural Language Processing and Social Humanities Computing Laboratory at Tsinghua University. The dataset contains approximately 840,000 news documents covering 14 mainstream news categories, such as sports, technology, entertainment, and finance.

[0125] In this experiment, we focused on analyzing small-sample learning scenarios with balanced datasets. For effective evaluation, we first randomly selected 100 and 1000 instances from each category in the ABDCD and THUCNews datasets, respectively, as the base training and test sets. Next, to investigate the impact of augmented data on model performance, we added new instances generated using a data augmentation model to these base instances. This approach allows for a systematic analysis and evaluation of the method's performance on training data of varying sizes, including both original and artificially generated samples.

[0126] Table 2

[0127]

[0128] 2. To evaluate the effectiveness of the data augmentation method, a series of downstream classification experiments were designed and conducted for the small sample generation task. In each experiment, class-balanced subsampling was performed from the training set to ensure an equal number of samples for each class, thereby addressing the imbalanced class distribution. Seven different data classes were selected as the original samples, and the data augmentation process was systematically applied to all subsamples, followed by downstream classification experiments. The entire process was repeated on each subsample to ensure the reliability and consistency of the results.

[0129] For the classification experiments, a base-size BERT was selected as the classification model. The BERT classifier model was initialized using pre-trained weights provided by the `bert-base-chinese` library. These weights were obtained through unsupervised learning on a large-scale Chinese corpus. The classification model was trained by introducing augmented data, aiming to capture the complex semantic and structural features of the input text from the deep bidirectional Transformer encoding output embeddings generated by the classification model, thereby achieving accurate prediction of class labels. This process not only validated the effectiveness of data augmentation strategies in improving model generalization and classification accuracy, but also demonstrated their potential for handling small-sample learning tasks.

[0130] Through in-depth analysis of classification experiment results, we aim to evaluate the specific impact of augmented datasets on classification model performance, thereby verifying the effectiveness of the data augmentation methods employed. In particular, we focus on whether the data augmentation methods significantly improve classification results and their improvement in small sample class recognition accuracy.

[0131] 3. Data enhancement model comparison experiment

[0132] To validate the effectiveness of the data augmentation method described in this paper, we used five typical data augmentation methods as baseline comparison models: EDA (Xie et al., 2020), back-translation, CBERT (Wu et al., 2019), GDA (KumarV et al., 2020), and AugGPT (Dai et al., 2023). We designed a series of classification experiments, augmenting two classification datasets using different data augmentation methods. The augmented data were then used alongside the original data to train downstream classification models. The representative BERT classification model was selected as the evaluation tool to ensure the broad applicability and reliability of the results.

[0133] Table 3 Data augmentation comparison experimental results

[0134]

[0135] For the open domain classification dataset THUCNews, the data augmentation method of the present invention has shown significant advantages in the BERT model. Specifically, after using the present invention for data augmentation, the classification accuracy of BERT reached 98.57%, compared with 92.58% without enhancement, and compared with other enhancement methods, it has also been significantly improved. This shows that the present invention can not only greatly improve the accuracy of classification tasks, but also effectively cope with the challenges in open domain text classification, and provide strong support for complex semantic understanding. In addition, the augmented data generated by the present invention helps the model better capture subtle differences in the text, thereby improving its generalization ability and robustness.

[0136] In the ABDCD dataset, although the pre-trained classification model itself already possesses a certain degree of generalization ability, its classification accuracy is only 85.14%. After introducing the data augmentation method of the present invention, the accuracy jumps to 94.86%. This result shows that the present invention can not only effectively capture domain-specific features, but also significantly improve the model's understanding and prediction capabilities for rare or complex labels, thus showing a huge advantage in classification tasks in professional fields. Especially for small sample datasets, the present invention effectively alleviates the problem of data scarcity by generating high-quality data, further improving the performance of the model.

[0137] A comprehensive analysis of the experimental results from the THUCNews and ABDCD datasets reveals the following conclusions: 1) Significantly improved classification accuracy: The present invention outperforms other data augmentation methods on both datasets, and the improvement is particularly pronounced when dealing with complex text classification tasks. Whether it is large-scale open-domain data or small-scale specialized domain data, the present invention can significantly improve the accuracy of the classification model; 2) Enhanced model generalization: The augmented data generated by the present invention not only increases the data volume but also enriches the data diversity, enabling the model to better capture subtle differences in the text, thereby improving its generalization and robustness.

[0138] 4. Experiment on uneven data distribution in aquaculture

[0139] In classification tasks, the category distribution of the data set has an important impact on the performance of the model. Especially when dealing with unbalanced data, the model tends to favor the majority class, resulting in a significant decrease in the prediction accuracy of the minority class. In order to evaluate the data enhancement method of the present invention in solving the problem of poor classification effect caused by uneven category distribution, an in-depth study was conducted in combination with source data distribution analysis and confusion matrix. A data distribution analysis was performed on the classification dataset ABDCD in the field of aquatic fishery disease prevention and control, and it was found that the category of "control measures" had a serious uneven distribution phenomenon in the original dataset. Specifically, the number of samples in this category is far less than that of other categories, accounting for only 2.1% of the total number of samples. This unbalanced distribution not only makes it difficult for the model to learn the characteristics of this category, but may also lead to overfitting or underfitting problems, thereby affecting the accuracy and robustness of the classification.

[0140] In order to more intuitively understand the impact of uneven data distribution on the classification model, a confusion matrix was constructed and a comparative analysis was performed at different stages before and after training. When using the original data for training, the classification confusion matrix is ​​as follows: Figure 3 As shown in the figure, due to the small number of samples in the category of "control measures", the model's prediction accuracy for this category is low and the misclassification rate is high, especially when confused with other categories. The model's predictions for most categories are relatively accurate, but the overall classification effect is affected by the drag of a few categories. In order to solve the above problem, the method of the present invention is used to augment the data of the "control measures" category. By generating high-quality augmented data, the present invention effectively increases the number of samples in this category, making its proportion in the training data more reasonable. After retraining the model, the confusion matrix is ​​constructed again, as shown in Figure 4 As shown, after the data augmentation of the present invention, the number of samples in this category has increased significantly, the model's prediction accuracy for this category has been greatly improved, and the misclassification rate has been greatly reduced. In particular, compared with before, the confusion between the control measures category and other categories has been significantly reduced, indicating that the model's ability to learn the features of this category has been significantly enhanced. With the improvement of the "control measures" category distribution, the classification performance of the entire model has also been comprehensively improved. The confusion matrix shows that the prediction accuracy of all categories has improved, especially the improvement of a few categories is the most significant. This not only proves the effectiveness of the present invention in solving the problem of uneven data distribution, but also demonstrates its positive impact on the overall model performance.

[0141] 6. Expanding Data Advantages

[0142] This experimental section focuses on intrinsic evaluation, comprehensively considering the impact of six different data augmentation methods on the Aquatic Biological Disease Dataset (ABDCD). These methods include EDA, BT, CBERT, GDA, AugGPT, and the method proposed in this paper. The performance of each method is measured using two key metrics: similarity (BERT Score) and richness (1-BLEU). Similarity reflects the similarity (i.e., fidelity) between the generated data and the original data, while richness measures the novelty and diversity of the generated text. Therefore, the quality of the augmented dataset is judged by combining these two metrics.

[0143] Table 4 Data augmentation comparison experiment

[0144]

[0145] As can be seen from Table 4, EDA and BT show the highest similarity, which are 96.72% and 96.56% respectively, which shows that these two methods can well preserve the semantic information of the original text. CBERT has the lowest similarity, which is 84.61%, which may be because it introduces more changes, resulting in semantic deviation. The similarity of GDA, AugGPT and the present invention is between the two, which are 90.51%, 94.13% and 94.95% respectively, showing that while they maintain a certain degree of similarity, they have also undergone a certain degree of change. With regard to richness (1-BLEU), the present invention shows the highest richness, with a score of 64.41%, which means that the text it generates has high novelty and diversity. CBERT follows closely behind, with a score of 58.59%, showing that it performs well in increasing text diversity. The richness of GDA and AugGPT are 45.19% and 41.88% respectively, indicating that they can also produce relatively rich text changes, but not as significant as the present invention and CBERT. The richness of EDA and BT is relatively low, at 14.81% and 29.75% respectively, indicating that these two methods mainly focus on minor changes in the text and fail to significantly improve the diversity of the text.

[0146] From a comprehensive perspective, the present invention provides relatively high richness (64.41%) while maintaining a high similarity (94.95%), which is a very good choice for application scenarios that need to retain original information while introducing new changes.

[0147] The above is a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A text data enhancement method for aquaculture disease prevention and control based on small model guidance, characterized by: The following steps are involved: S1. Construct the original sample set of aquatic biological diseases and perform fine-tuning on the corpus annotation to generate the aquatic biological disease annotation dataset; S2. Split each piece of data in the original sample set into a group of data segments using sentence-level segmentation; S3, build a Prompt framework that integrates prompt engineering and rule checking strategies; S4: Based on the aquatic biological disease annotated dataset constructed in S1, supervised fine-tuning is performed on the pre-trained large model to obtain the SFT domain large model; and the Prompt framework constructed in S3 is used to guide the SFT domain large model to generate enhanced data; S5. Use the small model as a supervisory model, input the slice data obtained in S2 and the enhanced data obtained in S4 into the supervisory model, and train the large model in the SFT field using the training method of the small model supervising the large model to obtain a large model for data augmentation in the field of aquatic diseases. The specific process is as follows: S5.

1. Construct a two-level supervision system in the supervision model, where one level focuses on the similarity evaluation between sentences and the other level focuses on the richness analysis of information; And the original slice data obtained by S2 Input into the supervised model to generate multi-objective reward scores, namely similarity scores and richness scores; The first Aquatic biological disease data; The original slice data Embed and augmented data Sentence Embedding Perform cosine distance calculation to get the similarity score : （7） （8） Richness score The calculation formula is as follows: （9） in, Represents Jaccard distance, used to describe and The difference between them, that is, richness; S5.2, corresponding to the two reward objectives of similarity score and richness score, two gating layers are introduced into the supervised model to comprehensively consider the balance between the two dimensions and their relationship with the enhanced data; specifically, the input and its threshold conditions ,as well as and its threshold conditions , the output of the gating layer Expressed as: （10） S5.

3. Binary labels generated by formula (10) 、Prompt word and enhanced data A dataset was constructed and the parameters of the large model in the SFT field were trained using the KTO reinforcement learning method to obtain a large model for data augmentation in the field of aquatic diseases. The loss function of the KTO training was defined as follows: （11） in: （12） （13） （14） （15） in, is the KL divergence term, It is derived from DPO reward Functional form; For the model being aligned, is the reference model, σ is the logarithmic function, and are the hyperparameters of the desired output and the undesirable output, respectively; The training parameters of the SFT domain large model are as follows: （16） in, is a hyperparameter, To adopt The model being aligned after LORA fine-tuning, is a trainable parameter; S6. Use the data augmentation model in the field of aquatic diseases obtained in S5 to generate enhanced data, and perform noise removal and aggregation processing on the enhanced data to obtain the final long text data.

2. The method for enhancing text data of aquaculture disease prevention and control based on small model guidance according to claim 1 is characterized in that: The specific process of S1 is as follows: S1.

1. Search using the name of aquatic biological disease as a keyword, collect information on several attribute types corresponding to each disease name, and form an original sample set The attribute types include disease type, pathogen / cause, target of harm, main symptoms, prevalence, treatment methods, control measures and prevention methods; S1.2, using the attribute-value pair data annotation method in structured annotation, the original sample set The aquatic biological disease data in the original sample set are annotated; Each aquatic biological disease data By core entity and its associated eight entity attributes Composition, among which the core entity Name of disease, entity attributes That is, the attribute type, , each entity attribute includes m Attribute description, so each aquatic biological disease data is represented as , which is expressed as ,in Represents labeled data. All aquatic biological disease data are labeled according to this logic to form a labeled dataset. .

3. The method for enhancing text data of aquaculture disease prevention and control based on small model guidance according to claim 2 is characterized in that: The specific method of S2 is: For the original sample set In a piece of aquatic biological disease data, first extract the core entity and related entity attributes ; Then the original data is segmented at the sentence level according to entity attributes to form a series of new data records ; each By a specific core entity The corresponding entity attributes This process is expressed as: （1） in, Indicates the operation of splitting the original data at the sentence level; each Indicates that from the original sample set The first i Sentence-level data segments.

4. The method for enhancing text data of aquaculture disease prevention and control based on small model guidance according to claim 3 is characterized in that: In the S3, the construction process of the Prompt framework includes: S3.1, adopt the framework of "purpose + requirements + format" to design is the final generated prompt word, is the goal of generating content, is the requirement for generating content, is the format of the output content, then the design process of the prompt word can be expressed by the formula: （2） in, The function represents the process of prompt word design, which accepts three parameters: target ,Require and format , and output an optimized prompt word ; S3.

2. After the content is generated, an automated mechanism is used to check whether the generated data meets the predetermined requirements. If it does not meet the initial set requirements, it will be regenerated until the threshold conditions are met. The entire process can be expressed as follows: （3） in, The function's workflow is as follows: （4） in, Indicates the final generated content; is the amount of content that needs to be generated; is the threshold condition, i.e., the maximum number of attempts; Indicates the Secondary generated content; Is a validation function that checks Whether the predetermined standards are met.

5. The method for enhancing text data of aquaculture disease prevention and control based on small model guidance according to claim 4 is characterized in that: The specific process of S4 is: The parameters of the pre-trained large model are adjusted using the aquatic biological disease annotation dataset. After fine-tuning with LORA supervision, the SFT domain model is as follows: （5） Using the Prompt framework to guide large models in the SFT field to generate enhanced data : （6） in, Indicates a prompt word.

6. The method for enhancing text data for aquaculture disease prevention and control based on small model guidance according to claim 5 is characterized in that: The pre-trained large model includes a ChatGLM model, an LLaMA model or a Qwen model.

7. The method for enhancing text data for aquaculture disease prevention and control based on small model guidance according to claim 5 is characterized in that: The specific process of S6 is as follows: S6.

1. Assume that the set of all data generated by the data augmentation model in the field of aquatic diseases is is a mapping function, for each data point , which generates a two-dimensional vector , representing the similarity score and richness score respectively, select the best layer by layer, dynamically adjust the selection ratio, and finally obtain high-quality data , the preliminary screening process is expressed as: （17） in, Represents the similarity screening index, represents the richness screening index; S6.2, in order to Restore to the original long text knowledge, using a random extraction and aggregation method: is the original sliced ​​dataset, where each element is a single sentence containing a single attribute; yes middle After data enhancement and processing, the data set is ,in , represents the number of data augmentations; for Augmented data of the same attribute in Random extraction is performed to form long text augmentation data. The specific implementation is as follows: （18） in, is an aggregate function, Indicates from Each Corresponding Randomly extract from the augmented versions and merge into one long text ; Generate long text augmentation data Amount of data .

8. An electronic device, characterized in that: It includes a memory, a processor, and a computer program stored in the memory and capable of running on the processor; when the processor executes the computer program, the electronic device executes the small model-guided aquaculture disease prevention and control text data enhancement method as described in any one of claims 1 to 7.

9. A storage medium, characterized in that: The storage medium includes a computer program, and when the computer program is run on an electronic device, the electronic device executes the aquaculture disease prevention and control text data enhancement method based on small model guidance as described in any one of claims 1 to 7.

Images