Plant diseases and pests are significant factors determining both the yield and quality of crops, which can be addressed by means of artificial intelligence (Spence et al., 2020). These diseases and pests represent a form of natural disasters that disrupt the healthy growth of plants, potentially leading to plant mortality throughout the entire development stage, from seed formation to seedling growth (Liu and Wang, 2021).

Traditional approaches to plant pest and disease classification have predominantly relied on manual inspections and specialized knowledge, which are labor-intensive, time-consuming, and prone to human mistakes and biases (Xing and Lee, 2022). With the rise of machine learning and computer vision, automated image-based classification methods have gained widespread attention for their potential to improve efficiency and accuracy (Domingues et al., 2022). For example, Shoaib et al. proposed advanced deep learning models for plant disease detection, highlighting the effectiveness of Convolutional Neural 85 Networks (CNNs) in learning hierarchical features from images (Shoaib et al., 2023). Some classical 86 architectures such as AlexNet, VGGNet, ResNet, and Inception have been employed to classify diseases in various crops, including tomato, rice, maize, and citrus (Sumaya et al., 2024). For instance, Yueteng et al. demonstrated that an improved ResNet architecture enhances recognition accuracy in complex plant disease datasets (Yueteng et al., 2021). Furthermore, traditional machine learning models like Support Vector Machines (SVM), k-Nearest Neighbors (k-NN), and Random Forests have been deployed, often in conjunction with manually extracted features such as color, texture, and shape descriptors. For example, Kale et al. analyzed crop disease detection using these classifiers and found that while effective under certain conditions, they often struggle with generalizability across diverse environmental conditions and are limited when dealing with visually similar symptoms among different diseases (Kale and Shitole, 2021). However, these methods often struggle with generalizability across diverse environmental conditions and are limited when dealing with visually similar symptoms among different diseases.

Recently, to overcome the limitations of single-modal approaches, there have already been some efforts on exploring multi-modal learning frameworks for plant pest and disease classification, which integrate heterogeneous data sources, such as images, textual descriptions, sensor data, and environmental metadata (Yang et al., 2021). This paradigm aims to enhance the robustness and contextual awareness of classification systems (Liu et al., 2025a). For example, Wei et al. proposed a multi-modal transformer architecture for citrus pests and diseases classification, where both image and text features are encoded and aligned through a cross-attention mechanism, enabling improved retrieval and identification performance (Wei et al., 2023). Similarly, Duan et al. introduced a multimodal system combining RGB images, text data, and environmental cues to facilitate pest detection and classification, demonstrating superior performance over image-only models, especially in complex agricultural scenarios (Duan et al., 2023). Wang et al. proposed Agri-LLaVA, an advanced multimodal assistant enriched with domain knowledge, designed specifically for managing 108 agricultural pests and diseases. Agri-LLaVA is trained on an extensive multimodal dataset, containing more than 221 varieties of pests and diseases, amounting to roughly 400,000 data samples. By integrating domain-specific knowledge into its training process, Agri-LLaVA demonstrates superior performance in both multimodal agricultural dialogue and visual comprehension, offering innovative solutions to tackle pest and disease challenges in agriculture (Wang et al., 2024b). These approaches leverage the complementarity of modalities, while images provide morphological cues, textual and contextual data supply semantic and environmental understanding, which proves to be useful for fine-grained and field-based classification tasks.

Recently, to overcome the limitations of single-modal approaches, there have already been some efforts on exploring multi-modal learning frameworks for plant pest and disease classification, which integrate heterogeneous data sources, such as images, textual descriptions, sensor data, and environmental metadata (Yang et al., 2021). This paradigm aims to enhance the robustness and contextual awareness of classification systems (Liu et al., 2025a). For example, Wei et al. proposed a multi-modal transformer architecture for citrus pests and diseases classification, where both image and text features are encoded and aligned through a cross-attention mechanism, enabling improved retrieval and identification performance (Wei et al., 2023). Similarly, Duan et al. introduced a multimodal system combining RGB images, text data, and environmental cues to facilitate pest detection and classification, demonstrating superior performance over image-only models, especially in complex agricultural scenarios (Duan et al., 2023). Wang et al. proposed Agri-LLaVA, an advanced multimodal assistant enriched with domain knowledge, designed specifically for managing 127 agricultural pests and diseases. Agri-LLaVA is trained on an extensive multimodal dataset, containing more than 221 varieties of pests and diseases, amounting to roughly 400,000 data samples. By integrating domain-specific knowledge into its training process, Agri-LLaVA demonstrates superior performance in both multimodal agricultural dialogue and visual comprehension, offering innovative solutions to tackle pest and disease challenges in agriculture (Wang et al., 2024b). In addition, Zhao et al. introduced PlanText, a gradually masked guidance framework to align image phenotypes with trait descriptions for plant disease texts, further highlighting the potential of integrating visual and textual modalities in plant health analysis (Zhao et al., 2024). Meanwhile, Dong et al. developed PlantPAD, a large-scale image phenomics platform for plant science, which provides high-quality resources for training and validating plant disease classification systems (Dong et al., 2024). These approaches leverage the complementarity of modalities, while images provide morphological cues, textual and contextual data supply semantic and environmental understanding, which proves to be useful for fine-grained and field-based classification tasks.

Although the above-mentioned multi-modal approaches have shown promise in plant pests and diseases classification, most existing methods primarily treat non-visual modalities as auxiliary inputs to enhance image-based features. This image-centric design often overlooks the independent value and discriminative power of other modalities, particularly textual data. In real-world agricultural scenarios, textual descriptions are typically colloquial, non-standard, and context-dependent, posing significant challenges to conventional multi-modal fusion strategies. While some studies have explored robust textual encoding techniques to handle noisy or weakly structured inputs (Zhu et al., 2023), these characteristics nevertheless result in a persistent semantic gap that current models struggle to bridge, thereby limiting their robustness and generalizability. These characteristics lead to a semantic gap that current models struggle to bridge, limiting their robustness and generalizability. To address these limitations, in this paper, we propose a novel approach to improve the model’s capacity to understand and utilize natural language expressions effectively for more accurate and practical plant disease classification.

While multimodal and text-based approaches have achieved progress, existing models still face significant challenges when processing colloquial, non-standard user inputs. Farmers’ symptom descriptions are often short, vague, and expressed in everyday language, which are inconsistent with the professional terminologies used in agricultural knowledge bases. For example, models trained on standardized datasets struggle to align colloquial expressions with technical disease terms, leading to misclassification or failure to recognize symptoms. Moreover, the semantic ambiguity and variability of colloquial text introduce additional noise, weakening the model’s ability to capture fine-grained distinctions across different disease categories. These limitations further underscore the necessity of developing methods that can effectively bridge colloquial language with domain-specific knowledge, which is precisely the problem our study seeks to address.

Prompt-tuning has surfaced as an efficient and effective method for adjusting Pre-trained Language Models (PLMs) to downstream tasks without requiring full model fine-tuning (Liu et al., 2021). This paradigm transferred downstream tasks through cloze-style objectives, which is particularly attractive in resource-constrained settings due to its efficiency and ability to preserve general language knowledge encoded in PLMs. The evolution of prompt-tuning includes both discrete and soft prompt methods. Early work in manual prompt design relied on human intuition to craft natural language prompts that could guide PLMs toward the desired behavior, including relation extraction (Han et al., 2021), knowledge probing (Petroni et al., 2019), and text classification (Hu et al., 2021). For example, Han et al. introduced a prompt-tuning model with rules for many-class classification tasks, encoding prior knowledge into prompt-tuning via logic rules and proposing manually designed sub-prompts to construct task-specific prompts (Han et al., 2021).

However, the manually created prompt proved to be inflexible and suboptimal, leading to the development of automated prompt generation strategies (Li and Liang, 2021). In the soft prompt-tuning, continuous embeddings are served as prompts and optimized while keeping the PLM’s weights frozen. For instance, Shin et al. developed the AUTOPROMPT method for generating prompts across various NLP downstream tasks (Shin et al., 2020). In the method, an auto-prompt consisted of the input sentence and the set of trigger tokens. These tokens remain consistent across all inputs and are determined via a gradient-based search mechanism. Wu et al. proposed an information-theoretic approach that framed soft prompt-tuning as optimizing the mutual information between the prompts and other model parameters (Wu et al., 2023). The technique involved two loss functions to achieve proper prompt initialization and extract relevant task-specific information from downstream tasks. Zhu et al. proposed a soft prompt-tuning method for short text stream classification (Zhu et al., 2025), which builds the verbalizer using internal knowledge rather than retrieving from external knowledge bases, further optimizing it through additional tailored strategies. Considering the advantages of soft prompt in contrast to manually crafted prompts, in this paper, we introduce the soft prompt-tuning method for colloquial descriptions in plant pest and disease classification.

As shown in Figure 1 , the proposed method first leverages the AgriBERT (Chen et al., 2024) for named entity recognition model to extract key agricultural entities from the obfuscated text, with relevant attribute information retrieved from the AgriKG (Chen et al., 2019) agricultural knowledge graph to effectively supplement domain knowledge. Then, the user-provided fuzzy description and the retrieved knowledge are concatenated to the soft prompt-tuning model. For the prompt-tuning method, the external verbalizer further enriches the model’s understanding by mapping informal expressions to corresponding technical terms using structured agricultural knowledge, allowing the model to interpret and classify colloquial symptom descriptions more accurately. By leveraging the generalization capabilities of PLMs and integrating domain knowledge through the constructed verbalizer, our approach notably enhanced both the precision and interpretability of pest and disease predictions.

First, we utilize the AgriBERT named entity recognition model, specifically trained for agricultural texts, to extract relevant entities from the input fuzzy agricultural text. This model captures contextual information through a multi-layer bidirectional self-attention mechanism and incorporates a global pointer mechanism for entity localization.

For the input fuzzy text T f u z z y = ( t 1 , t 2 , … , t n ) , where t_i represents the i-th word in the text, the AgriBERT model outputs a set of entity labels E = { e 1 , e 2 , … , e m } , where each entity e_i includes the entity type and its position within the text. However, this positional information is still relatively coarse and cannot guarantee precise boundary detection. We use the global pointer mechanism P( e i ) to represent the specific position of entity e i in the text, as described by the following formula (Equation 1):

where start_i and end_i represent the start and end positions of entity e_i , respectively. Thus, e i provides the semantic label, while P( e i ) precisely anchors the boundary, thereby enhancing the model’s robustness in handling overlapping or ambiguous entities.

After extracting the entities, we leverage AgriKG, a publicly available agricultural knowledge graph, to query domain-specific knowledge related to the entities. By querying AgriKG, we obtain the corresponding relevant knowledge K_i for each entity e i , which contains various attributes related to the entity. Let K_i represent the set of knowledge fragments obtained from AgriKG. We organize these knowledge fragments as follows (Equation 2):

where each knowledge fragment k j provides specific information relevant to entity e i and contains domain related knowledge in agriculture.

Next, we concatenate the user-provided fuzzy description T f u z z y = ( t 1 , t 2 , … , t k ) and the relevant knowledge fragments K_i retrieved from AgriKG. The fuzzy text T f u z z y represents the words in the non-expert language provided by the user. We concatenate the user’s description T f u z z y with the knowledge fragments from AgriKG in the following format (Equation 3):

where [SEP] is a separator used to distinguish the original description from the knowledge fragments. The concatenated text contains both the user’s non-expert description and the supplemental domain knowledge, thereby enhancing the professionalism and completeness of the text.

To further illustrate this process, we provide two running examples in Table 1 . Each input is first parsed by AgriBERT-NER to extract entities, then enriched with compact knowledge snippets from AgriKG, and finally concatenated into the enhanced description. As shown in the table, the pipeline effectively aligns colloquial farmer expressions with formal agronomic terminology, leading to accurate classification results.

We adopt AgriBERT as the backbone PLM. This model is specifically trained on agricultural text tasks, enabling strong capabilities in agricultural terminology recognition and semantic representation. Essentially, AgriBERT follows the BERT architecture, consisting of 12 Transformer encoder layers, each with a 768-dimensional hidden representation and 12 self-attention heads.

In contrast to prompt-tuning methods relying on manually crafted templates, our method utilizes soft templates learned within a continuously optimized prompt space. When integrated with the enhanced description x e n described earlier, this approach enables more adaptive text recognition by the model, and can be formulated as (Equation 4):

where x_en represents the enhanced description obtained by concatenating the fuzzy text with the knowledge fragments introduced in Section 3.2, u i denotes the i t h learnable token, the prompt T is subsequently passed through the encoder of aPLM to generate hidden states h i , … , h x e n , … , h n , h MASK . Accordingly, the soft prompt is formulated as (Equation 5):

To further enhance the model’s ability to capture temporal and contextual information in non-standard expressions, we integrate a two-layer bidirectional LSTM encoder head into the prompt-learning framework. The input size, hidden size, and embedding size are all set to 768 to maintain consistency between forward and backward information flows. During inter-layer representation transfer, each LSTM layer employs the standard hidden-state propagation mechanism to preserve information progression. The final output is obtained by concatenating the hidden states from both directions, which is then fed into the classifier for prediction. This process can be visualized as (Equation 6):

where h_i denotes the hidden state input at the i-th position of the input sequence, derived from the encoder of AgriBERT. Specifically, it represents the contextualized embedding obtained by feeding the soft prompt together with the textual input into the Transformer architecture. The final fused representation h i ' is formed by concatenating the hidden states from the forward and backward directions of the BiLSTM, thereby capturing bidirectional contextual information.

Ultimately, the model improves its performance and output quality by determining the optimal values of variables via the loss function, as illustrated in (Equation 7).

To adjust the model parameters, we adopt the cross-entropy loss, which quantifies the divergence between predicted outputs and ground-truth labels. The objective function is formulated as (Equation 8):

where N denotes the number of training instances, y * is the gold label, and | θ | 2 represents the L2 penalty on parameters θ . The penalty term helps alleviate overfitting by restricting parameter magnitudes, and the coefficient α balances the impact of regularization in the overall loss.

Prompt-tuning studies (Schick and Schutze, 2020) have shown that aligning label words with their target categories y helps reduce the mismatch between textual input and label representation. This process, referred to as automatic label word selection (Gao et al., 2020) or verbalization (Schick et al., 2020), can be formally expressed as (Equation 9):

where v i represented a word in the verbalizer. In our method, we build the verbalizer by leveraging words extracted from an external knowledge graph. This method expands semantic diversity while promoting greater robustness and generalizability of the verbalizer.

To retrieve mapping words associated with the target categories from a knowledge graph, we employ Related Words ¹ as our external source. This knowledge graph aggregates multiple resources, such as word vectors, ConceptNet (Speer et al., 2017), and WordNet (Pedersen et al., 2004), allowing us to extract an initial set of words v for each category label y, thereby constructing the base verbalizer. Given the vast amount of text and the possibility of noise or irrelevant content, we implement three optimization strategies to refine the extracted words. These strategies are designed to capture various facets of the expanded word characteristics, aiming to uncover the underlying intent of the original text. The specific methods are outlined below:

FastText Similarity: A commonly employed method for improving verbalizer construction consists of evaluating the semantic similarity between category labels and their extended label terms. This approach uses the FastText embedding model to generate vector representations and calculate the cosine similarity between category label terms and their expanded equivalents. Let E y and E v denote the embeddings of a category label y and an extended label term v , respectively. The cosine similarity is expressed as (Equation 10):

where g refers to the dimension of the word embedding, while E y i indicates the i t h component of the vector E y .

Notably, to ensure relevance, only the top N expanded words with the highest cosine similarity to the category label are preserved, while those with low similarity are excluded.

Probability Prediction: Leveraging contextual cues and prompt templates to estimate the likelihood of masked tokens is a crucial technique in refining the verbalizer. This is achieved using a PLM(e.g., BERT), which outputs the probability distribution over potential words filling the [MASK] position.

More concretely, given a prompt template T, the model masks certain words in the input and computes p(T[MASK]), the probability distribution over possible replacements. This distribution reflects the strength of association between each candidate word and the target category.

We apply BERT to obtain this distribution and select the top N most probable terms to expand the label word set.

Context Information: In order to enrich the label word set and effectively utilize the surrounding context of masked tokens, we propose an expansion strategy based on context windows. Rather than relying on conventional N-gram models, our approach leverages non-autoregressive PLMs like BERT to capture contextual dependencies. Given that BERT cannot directly estimate full-sentence generation probabilities, we address this constraint through the application of a symmetric sliding window approach.

Assuming a window size of c , the context centered around the [MASK] token can be represented as (Equation 11):

Within this framework, each word w i in the window is sequentially masked and input into the BERT model for calculating the loss associated with predicting the masked word (Equation 12):

where V denotes the vocabulary set, 1 is the indicator function, and p ( v i = w i | W w i ) represents the predicted probability distribution of BERT conditioned on W with w i excluded.

In the experiments, label word candidates are sorted according to their sequence loss L(W), and those with higher loss values are discarded. Only the words with the lowest losses are preserved. A fixed window size of c = 5 is used, and for consistency, the N = 15 identified by each of the three strategies are selected to construct the expanded label word set.

The combination of FastText Similarity, Probability Prediction, and Context Information enables a multi311 faceted enhancement of the verbalizer, thereby substantially improving the model’s semantic understanding of category labels.

Once the external knowledge-based verbalizer has been refined using the three proposed strategies, we compute the prediction score using a weighted average of the label word scores. In particular, the final prediction y ^ is obtained by aggregating the scores of all candidate categories according to their respective word weights. These weights are calculated based on the contribution of each word, as formulated below (Equation 13):

Here, V refers to the collection of label words linked to the category y , while | V y | indicates the size of this set. The probability function p ( [ MASK ] = v | x en ) evaluates how likely the label word v is, conditioned on the enhanced description x e n .

In this study, we use two benchmark English datasets: the PlantWild dataset and the GojiPest dataset.

PlantWild: PlantWild is a large-scale dataset for wild plant disease recognition, covering multiple healthy plant categories and plant disease categories. It contains over 50,000 images, with each plant disease category accompanied by rich textual descriptions. In this study, we primarily use the textual descriptions from this dataset, which provide detailed explanations of the fine-grained features of various plant diseases, helping the model identify subtle differences between them.

GojiPest: GojiPest is a cross-modal image-text dataset focused on goji plant pests and diseases. It supports tasks such as image collection, text creation, data augmentation, classification, and image-text pairing. The dataset includes images and textual descriptions for various common goji pests and diseases. Similar to the PlantWild dataset, we only use the textual descriptions from this dataset in our study, focusing on utilizing the descriptive information for enhancing the understanding and classification of goji plant pests and diseases.

In order to assess the effectiveness of our approach, we conducted comparisons with SOTA methods.

Stacked Denoising Autoencoders (SDA) (Zhu et al., 2019): SDA is a conventional unsupervised deep learning model that generates detailed feature representations for both the training and test datasets via an autoencoder. Subsequently, a classifier is trained on labeled data from the training set to carry out classification tasks on the test data.

TextCNN (Kim, 2014): A deep learning network that incorporates a convolutional layer to extract contextual features from labeled training data. Once trained, the model is applied to execute classification tasks on the test set.

BERT (Devlin et al., 2018): BERT (Bidirectional Encoder Representations from Transformers) is based on the Transformer framework. It reformulates tasks into cloze-style (fill-in-the-blank) questions, making it a robust baseline approach for a variety of NLP tasks.

AgriBERT (Chen et al., 2024): AgriBERT is a pre-trained language model specifically designed for agricultural domain texts. It is trained on a large corpus of agricultural literature and technical reports, making it more adept at understanding agricultural terminologies and contexts compared to general-domain PLMs like BERT.

Prompt Learning (PL) (Liu et al., 2021): This method integrates input data from each chunk into a pre-designed template, using only the category name to build the verbalizer in regular prompt-tuning. For consistency, the templates in PL align with those used in our experiments.

P-tuning (Liu et al., 2022): P-tuning is a method for soft prompt-tuning that involves learning continuous prompts by embedding trainable variables into the input representations, instead of using hand-crafted templates.

Mistral (Jiang et al., 2023): Mistral is an emerging large-scale language model created by the Mistral AI team. It is particularly known for its high computational efficiency and robust generative capabilities, outperforming similar models, especially in multimodal tasks.

LLaMA3 (Touvron et al., 2023): LLaMA3 is an efficient, large-scale language model developed by Meta, designed to function well in low-resource environments. It reduces the number of parameters to minimize computational costs while maintaining solid performance in natural language reasoning and generation tasks.

SimSTC (Liu et al., 2025b): A straightforward framework for graph contrastive learning applied to short text classification. The method performs graph learning on various component graphs related to text, generating multi-view text embeddings, upon which contrastive learning is directly applied.

In this experiment, to generate vague descriptions, we constructed an “expert-non-expert” parallel corpus and fine-tuned a large language model using LoRA technology, enabling it to convert specialized terms into more accessible expressions. Specifically, we first collected approximately 1,200 expert-level symptom descriptions from agricultural manuals and online agricultural knowledge bases. Each description was then paraphrased by GPT-4 into a farmer-style colloquial version, resulting in an expert–non-expert pair. Based on this corpus, we fine-tuned LLaMA-8B with LoRA, using a configuration of rank = 8, α = 16, learning rate = 2e-4, and 5 training epochs on dual NVIDIA A6000 GPUs. To ensure quality, we manually filtered the generated sentences to remove incomplete or duplicate expressions. Furthermore, three agricultural experts validated a random sample of 300 pairs, achieving an agreement rate above 90%. This process ensured that the constructed parallel corpus is reliable and suitable for subsequent fuzzification.

We then used the fine-tuned model to apply fuzzification to the text in the dataset. Next, we applied our method to enhance the fuzzified text, resulting in the final dataset for subsequent experiments. The dataset was separated into subsets for training, testing, and validation, with 70% of the data assigned to training, 20% designated for testing, and the remaining 10% kept for validation.

For methods based on deep neural networks and fine-tuned pre-trained language models (e.g., SDA, TextCNN, SimSTC, BERT and AgriBERT), the full training dataset was utilized, as these models necessitate large quantities of data for effective learning. Furthermore, we applied the hyperparameters specified in the original papers to maintain consistency and achieve optimal performance. For the prompt-tuning approaches (PL and P-tuning), a 20-shot configuration was implemented for both methods. To ensure fairness, we kept the parameter settings identical across these approaches: dropout rate was set to 0.5, learning rate to 3e-5, batch size to 32, and weight decay to 1e-5. The hyperparameters, including batch size and learning rate, were determined through repeated empirical experiments to achieve optimal performance. The models were trained for 5 epochs to guarantee thorough training and stable outcomes, with the Adam optimizer employed for parameter tuning. For large language models (e.g., LLaMA3 and Mistral), classification was performed directly by formulating prompts in a question-answer format. Unlike traditional models, these systems do not rely on conventional training techniques but instead fine-tune the prompts specifically for classification tasks.

The effectiveness of our methods is evaluated using the following four key metrics: As Equations 14–17 is shown below.

Accuracy (Acc): The proportion of correctly predicted samples compared to the total number of samples.

Precision (Pre): The ratio of positive samples among the predicted positive samples.

Recall (Rec): The ratio of correctly predicted positive samples to the total number of actual positive samples.

F1 Score (F1-S): The harmonic mean of Precision and Recall, used as a comprehensive measure of classification performance. A higher F1 score indicates better overall performance.

All experiments were conducted on a server equipped with an NVIDIA A100 GPU, a 64-core AMD EPYC 7763 processor, and 512 GB of memory. The experiments were performed using Python 3.9.16 and PyTorch 1.12.0 with CUDA support.

Table 2 presents the performance results of our method and baseline models across two datasets (PlantWild and GojiPest). Based on these experimental results, the following insights have been observed:

In conclusion, our method consistently outperforms baseline models across all subtasks, especially in few-shot learning and in handling noisy short-text descriptions. This highlights the advantages of prompt-based learning and knowledge expansion. Future work could explore integrating domain-specific knowledge to optimize prompt templates further and improve the model’s generalization abilities for real-world applications.

To comprehensively evaluate the robustness of our proposed method, we conducted additional experiments along three dimensions: comparisons with fine-tuned large language models, parameter-efficient fine-tuning, and advanced prompting strategies. Specifically, we systematically compared our text-based approach with fine-tuned Qwen, LLaMA with LoRA-based fine-tuning, and Qwen variants equipped with Chain-of-Thought (CoT) reasoning and Self-Consistency (SC) voting under the same few-shot setting.

The experimental results, as shown in Table 3 , reveal three major findings: (1) Directly using LLMs to analyze agricultural text lacks in-depth reasoning, resulting in performance inferior to methods that incorporate CoT and SC; (2) In scenarios with limited data, LoRA struggles to fully capture the diversity and complexity of the task, making it difficult to generalize effectively to different agricultural text scenarios. This lack of sufficient data support during fine-tuning leads to unstable model performance, with significant drops in certain tasks. (3) Advanced prompting strategies (CoT + SC) bring improvements in some scenarios, but their overall performance is unstable and remains lower than the experimental results of our method.

In addition, an in-depth error analysis shows that LLMs tend to underperform on datasets containing short, ambiguous, and highly colloquial expressions, such as the Tomato dataset. This inconsistency can be attributed to its sensitivity to informal expressions and its limited ability to generalize across heterogeneous agricultural data. In contrast, our method leverages knowledge-enhanced soft prompt-tuning to explicitly bridge colloquial farmer descriptions with formal agricultural terminology, thereby achieving more stable and reliable performance even under noisy and diverse input conditions.

In this experiment, the design of templates was pivotal in influencing the model’s performance. To evaluate the effect of various hand-crafted and soft templates on classification tasks related to plant diseaseand pest descriptions, we created and tested several templates, the specifics of which are outlined in Table 6 . These tasks involve complex short-text descriptions, requiring the model to extract meaningful features and information based on the guidance provided by the templates.

The experimental results, displayed in Table 7 , demonstrate that specific hand-crafted templates successfully direct the model in grasping the essential elements of the tasks, especially in tasks like Insect2, where the model effectively identifies the key features of pest descriptions. However, as the complexity of the tasks and the diversity of the datasets increase, the limitations of using a single, fixed template become evident. For instance, vague or incomplete descriptions may hinder hand-crafted templates from fully leveraging the potential of the data. Consequently, we introduced soft template generation in our approach to improve the model’s ability to process uncertain and ambiguous text. The experimental results demonstrate that, even with limited training data, soft templates can be precisely adapted to the data, thereby creating an optimal theoretical prompt that considerably boosts classification accuracy.

We conducted further experiments to evaluate the effect of different hyperparameters, such as learning rate and batch size, on the experimental results. The findings are presented in Figure 4 . The learning rate governs the magnitude of parameter adjustments during model training. Based on the experimental results, the best performance across most tasks was obtained with a learning rate of 3e-5. This indicates that, within a specific range, a higher learning rate can expedite model convergence and assist the model in adapting to variations in the data. However, for some tasks, setting the learning rate too high can lead to instability during training, emphasizing the importance of fine-tuning the learning rate based on the specific requirements of each task.

In addition to learning rate, batch size is another critical hyper-parameters that influences training dynamics, memory usage, and model convergence. Our findings show that the model performed optimally with a batch size of 32. This suggests that, within an appropriate range, smaller batch sizes can facilitate faster convergence, helping the model adapt quickly to training data, especially for tasks with more complex patterns. On the other hand, larger batch sizes contribute to more stable training and enhanced performance in certain tasks. These results indicate that different datasets and tasks may require distinct batch sizes to achieve the best performance.

Furthermore, we assessed the impact of training epochs on model performance, as shown in Table 8 .

The results indicate that while model performance improves with more epochs, excessive training can lead to diminishing returns. Specifically, in the Apple task from the PlantWild dataset, the model achieved an optimal accuracy of 95.19% after 15 epochs, with a slight decrease to 94.23% at 20 epochs. Similar trends were observed for the Corn and Cucumber tasks, where the best performance was also achieved at 15 epochs, with no significant gain at 20 epochs. These findings suggest that a moderate number of epochs promotes model convergence and generalization, while too many epochs may lead to overfitting. This pattern was further confirmed in the Insect Pest dataset, where 15 epochs also yielded the best results in both Insect1 and Insect2 tasks. In conclusion, selecting an appropriate number of epochs is crucial for optimizing model performance, with 15 epochs proving to be ideal for most tasks.

In this section, we constructed a small-sample colloquial description dataset by collecting Q&A pairs from the agricultural platform “Ask Extension ² ” and carefully curating them through filtering and grained annotation. The dataset contains 1,000 samples, covering five categories of plant diseases and pests: Maize Leaf Spot, Rice Blast, Rice Planthopper, and Wheat Powdery Mildew.

On this dataset, we conducted validation experiments using fine-tuned PLMs (AgriBERT), prompt-based learning methods (PL and P-Tuning), the LLM Qwen, as well as our proposed method. The experimental settings and parameters were kept consistent with those in the main experiments. The results are reported in Table 9 . It can be observed that our method outperforms all other approaches across Accuracy, Precision, Recall, and F1, achieving an F1 Score of 79.70%, which is significantly higher than the second-best method, AgriBERT. In contrast, PL and P-Tuning perform poorly on this small-sample colloquial dataset, indicating the limited generalization capability of traditional prompt-based learning. Although Qwen performs better than the prompt-based learning methods, its adaptation to colloquial data is still insufficient without fine-tuning. Overall, our method demonstrates clear advantages in handling colloquial and standard expressions in classification tasks, providing more accurate and stable predictions across all sample categories.