Multi-perspective prompt fusion for zero-shot classification of agricultural news texts with large language models Authors: Qiaoyun Liu, Yaoyao Ding, Hongyan Liu, Zeling Xu, Yuntao Zou Published: 2026 ABSTRACT AbstractIn the era of smart agriculture, agricultural information and related event news spread rapidly across social media and online platforms. Efficient classification of such massive agricultural texts holds practical value for optimizing production decisions and enhancing risk early warning. The wide dissemination of these texts also raises the demand for accurate and timely content analysis. However, the domain vocabulary sparsity, semantic ambiguity, and annotated data scarcity of agricultural texts pose notable challenges for classification tasks. Existing zero-shot classification methods based on large language models predominantly design prompt templates from a single perspective, making it difficult to simultaneously cover multiple levels of text analysis, while classification judgments also lack cross-validation mechanisms from multi-dimensional information. To address these issues, this paper proposes a multi-perspective prompt fusion framework that constructs multiple complementary analytical perspectives based on hierarchical differences in text analysis, forming multi-level coverage from local terminology features to overall thematic attribution, and guiding large language models to perform classification judgment from different cognitive dimensions. This paper further constructs a dual-dimensional consistency measurement mechanism integrating decision-level and confidence-level metrics, and employs an adaptive threshold-driven disagreement detection and arbitration mechanism to conduct secondary judgment on low-consistency samples. Experiments on the public datasets PestObserver-France and HumAID demonstrate that the proposed method achieves classification accuracy of 85.3% and 68.5%, corresponding to improvements of 3.2% and 3.7% over the strongest baseline ChatAgri. The expected calibration error decreases from 0.135 and 0.148 to 0.098 and 0.112, representing reductions of 27% and 24% respectively. Perspective heterogeneity analysis shows that the pairwise prediction agreement rate among the five perspectives is 18 to 21 percentage points lower than that of the five-path sampling of Self-Consistency, confirming the contribution of heterogeneous perspective design to classification robustness. The framework currently relies on text input only and provides a basis for extension to multimodal agricultural information.Keywords: Zero-shot learning, Large language models, Agricultural text classification, Information dissemination, News text miningSubject terms: Education, Mathematics and computing Introduction In the era of smart agriculture, massive volumes of agricultural text information continue to accumulate1. Texts such as agricultural news, policy bulletins, pest and disease reports, and market dynamics carry critical knowledge including crop production guidance, disaster early warning, and market analysis. Efficient acquisition of this knowledge holds significant theoretical value and pressing practical demand for optimizing agricultural production decisions, supporting policy formulation, and enhancing risk management. Whether the domain knowledge embedded in agricultural texts can be effectively acquired and utilized directly affects the effectiveness of smart agriculture informatization system construction. Insufficient knowledge utilization efficiency will cause critical information to be submerged in massive texts, subsequently leading to cascading decision-making failures such as delayed pest and disease warnings and lagging market risk assessments. However, the domain specificity of agricultural texts poses significant challenges for efficient classification. Agricultural texts involve the interdisciplinary use of specialized terminology from agronomy, plant pathology, agricultural economics, and meteorology, exhibiting significant domain vocabulary sparsity, cross-lingual expression variability, and annotated data scarcity. Semantic-level complexity is equally prominent. The semantic ambiguity of specialized terminology, the implicitness of sentiment expression, and the complexity of contextual dependencies are superimposed upon one another–the same term may point to entirely different category memberships in different contexts. This multi-dimensional semantic characteristic poses fundamental difficulties for classification modeling. Methods based on fine-tuning pre-trained language models such as BERT and RoBERTa have achieved significant progress in text classification tasks2,3, with related work achieving considerable improvements over traditional machine learning methods on agricultural text classification tasks. However, these methods share common limitations. The main bottleneck lies in their dependence on large-scale domain-annotated data, while annotation costs in the agricultural domain are high and professional knowledge barriers are substantial. Limited cross-lingual transfer capability and high deployment costs of large models further exacerbate application difficulties. These constraints restrict the scalable application of the fine-tuning paradigm in agricultural informatization. To address the fine-tuning paradigm’s dependence on annotated data, the academic community has begun exploring zero-shot classification paradigms based on large language models4. Large language models such as ChatGPT and GPT-4, leveraging their deep semantic understanding capabilities acquired from massive corpora and powerful instruction-following abilities, have demonstrated significant potential in zero-shot text classification tasks5,6. Related research indicates that zero-shot strategies can achieve or even exceed the classification accuracy of fine-tuning methods under specific conditions7,8, providing new possibilities for agricultural text classification without domain-specific fine-tuning. However, existing applications of large language models to agricultural text classification still exhibit several shortcomings. Mainstream methods design prompt templates based on a single classification perspective, either emphasizing semantic matching of category definitions or focusing on step-by-step reasoning chains. Text analysis typically involves multiple levels including lexical recognition, semantic understanding, sentiment judgment, contextual disambiguation, and topic induction, with information captured at each level possessing certain complementarity. A single perspective cannot simultaneously cover information from these different levels, and multi-dimensional information in agricultural texts, such as the domain indicativeness of specialized terminology, the category relevance of semantic connotations, and the implicit expression of sentiment orientation–may not be fully captured. Furthermore, single-perspective strategies leave classification judgments without cross-validation from multiple dimensions. When a particular perspective produces misjudgments due to variations in text feature distributions, the model cannot correct itself through complementary perspectives. To address these issues, this paper conducts improvement research based on existing zero-shot agricultural text classification frameworks and multi-path consistency voting methods. ChatAgri designed four strategies–manual prompts, model-generated prompts, zero-shot similarity prompts, and chain-of-thought prompts–achieving performance comparable to PLM fine-tuning methods without domain-specific fine-tuning9,10. Self-Consistency achieved significant improvements in complex reasoning tasks by sampling multiple reasoning paths and selecting the most consistent answer through majority voting11,12. However, each of these methods has shortcomings. ChatAgri’s four prompt strategies are essentially all based on a single classification perspective, lacking mechanisms for comprehensive judgment from multiple complementary levels. The diversity in Self-Consistency originates from random perturbations introduced by temperature sampling. Preliminary experiments show that multiple reasoning paths generated through temperature sampling exhibit high correlation, with an average pairwise prediction agreement rate of approximately 75%. Each path may make the same errors on similar samples, and this homogeneous diversity has limitations in achieving perspective complementarity. Based on the above analysis, the problem addressed in this paper consists of two aspects. First, how to design a multi-perspective prompting mechanism with hierarchical complementarity, enabling large language models to analyze the same agricultural text from different cognitive dimensions including lexical, semantic, sentiment, contextual, and thematic levels, so as to overcome the limited information coverage of single-perspective prompts. Second, how to construct a reliable fusion and arbitration mechanism on top of multi-perspective judgments that quantitatively evaluates both the decision consensus and the confidence level among perspectives and handles low-consistency samples in a targeted manner, so that classification results are improved not only in accuracy but also in quantifiable reliability. To address these shortcomings, this paper systematically investigates the application of multi-perspective prompt fusion strategies in zero-shot agricultural text classification tasks. At the multi-perspective prompt design level, this paper constructs five complementary analytical perspectives based on hierarchical differences in text analysis: the lexical perspective focuses on terminology feature recognition, the semantic perspective focuses on holistic meaning representation, the sentiment perspective focuses on attitude and stance analysis, the contextual perspective focuses on disambiguation constraint information, and the topic perspective focuses on macro-level topic attribution. These five levels form complementary coverage from local to global13,14. At the perspective consistency fusion level, this paper constructs a dual-dimensional consistency measurement mechanism integrating decision-level and confidence-level metrics. Voting consistency characterizes the degree of consensus among perspectives on category decisions, while confidence consistency characterizes the certainty degree of predictions. Both collaboratively assess the reliability of classification results. At the disagreement detection and arbitration level, this paper designs a disagreement detection mechanism based on adaptive thresholds, triggering secondary comprehensive judgment for low-consistency samples. This paper conducts comprehensive comparative experiments and ablation analyses on multiple agricultural text datasets, validating the contribution of each perspective and the complementarity among perspectives, aiming to provide methodological reference and practical guidance for zero-shot text classification in the agricultural domain. Related work Agricultural text classification methods Agricultural text classification research has undergone a paradigm shift from traditional machine learning to deep learning. Early studies primarily relied on handcrafted feature engineering, employing classifiers such as Support Vector Machines and Naive Bayes for categorizing agricultural news and pest reports15. These methods depend on feature templates designed by domain experts, resulting in limited generalization capability and difficulty in capturing deep semantic information. With the emergence of pre-trained language models, fine-tuning approaches based on BERT2, RoBERTa3, and similar models have become the mainstream paradigm, achieving significant performance improvements on agricultural text classification tasks16. However, the fine-tuning paradigm exhibits common limitations. The main bottleneck lies in its dependence on large-scale domain-specific annotated data. Agricultural texts involve multi-disciplinary terminology spanning agronomy, plant pathology, and agricultural economics, requiring experts with corresponding domain knowledge to participate in the annotation process, which is costly and efficiency-constrained17,18. Insufficient cross-lingual transfer capability and high deployment costs of large models further restrict the scalable application of the fine-tuning paradigm in agricultural informatization scenarios19. These limitations have prompted researchers to explore classification paradigms that do not require domain-annotated data. Zero-shot classification with large language models The emergent capabilities of large language models have provided a new avenue for zero-shot learning in text classification7. Models such as ChatGPT and GPT-4, leveraging their deep semantic understanding abilities acquired from massive corpora and strong instruction-following capabilities4, can accomplish classification tasks without observing any labeled samples6,20. Researchers have designed task-descriptive prompts to guide models in understanding classification intent, validating the effectiveness of zero-shot strategies on multiple general-domain datasets13. ChatAgri introduced this paradigm into the agricultural domain, designing four strategies: manual prompts, model-generated prompts, zero-shot similarity prompts, and chain-of-thought prompts9. This work explored the application potential of large language models in cross-lingual agricultural text classification, achieving performance comparable to fine-tuned pre-trained models without domain-specific fine-tuning. While the zero-shot paradigm effectively addresses the annotation dependency problem, existing methods still exhibit several limitations. Mainstream approaches design prompt templates based on a single classification perspective, either emphasizing semantic matching of category definitions or focusing on step-by-step reasoning chains21. The multi-dimensional semantic features of agricultural texts are difficult to fully capture through a single perspective, and classification judgments lack cross-validation mechanisms from multiple dimensions. When a particular perspective produces systematic misjudgments due to variations in text feature distributions or prompt design biases, the model cannot identify such biases nor self-correct. Multi-path reasoning and consistency aggregation To enhance the robustness of large language model reasoning, researchers have proposed multi-path reasoning and consistency aggregation strategies. Self-Consistency generates multiple reasoning paths through temperature sampling and selects the most consistent answer via majority voting, achieving significant performance improvements in mathematical reasoning and commonsense question-answering tasks11,22. The underlying assumption of this method is that complex reasoning tasks typically have multiple paths leading to the correct answer, and sampling multiple paths while aggregating results can effectively reduce random errors from single inference23. Multi-Perspective Self-Consistency further introduced multi-perspective thinking into code generation tasks, constructing consistency graphs from three perspectives–solution, specification, and test cases–and achieved excellent results24. However, existing multi-path methods have limitations in their sources of diversity. Self-Consistency relies on random perturbations introduced by temperature sampling; while the generated reasoning paths differ in expression, they share essentially the same analytical perspective, with each path approaching the problem from the same semantic dimension25. This homogeneous diversity fails to achieve genuine complementary verification across heterogeneous perspectives. When a systematic bias exists in a particular analytical dimension, multiple paths based on that dimension will simultaneously produce misjudgments. These limitations indicate that the effectiveness of multi-path methods depends not only on the number of paths but more critically on the degree of fundamental differences in analytical perspectives among paths26. Problem definition Zero-shot agricultural text classification aims to automatically categorize agricultural texts into predefined categories without requiring domain-specific annotated data. Let denote the collection of agricultural texts, where any text represents unstructured text to be classified, such as agricultural news, policy bulletins, or pest and disease reports. Let denote the set of candidate categories, where K represents the total number of categories, and each category corresponds to a specific agricultural topic or semantic domain. The objective of zero-shot classification is to learn a mapping function such that for any input text x, the model can predict its category without having observed any labeled samples of that category. This paper employs a multi-perspective prompt fusion strategy to realize the aforementioned mapping. Let denote the set of analytical perspectives, where M represents the total number of perspectives. For a given text x and perspective , the large language model generates a classification response according to the perspective-specific prompt template . After structured parsing, this response yields the category prediction and confidence estimate of perspective for text x. Based on the prediction results from M perspectives, the fusion function g integrates the judgments from all perspectives to obtain the final classification result. The formal representation of this process is: To support disagreement detection and arbitration mechanisms, this paper further defines a perspective consistency metric η to quantify the degree of agreement among multi-perspective predictions. When η falls below a preset threshold τ, the system triggers secondary arbitration or outputs an uncertainty warning. The global symbol definitions are presented in Table 1. Table 1Global symbol definitions.SymbolDefinitionSymbolDefinitionAgricultural text collectionSubset of perspectives predicting the winning categoryxSingle agricultural text to be classifiedVoting consistency metricCandidate category setConfidence consistency metricThe k-th candidate categoryηComposite consistency metricKTotal number of candidate categoriesBaseline consistency thresholdAnalytical perspective setτAdaptive consistency thresholdThe m-th analytical perspectivegMulti-perspective fusion functionMTotal number of analytical perspectivesFinal classification predictionLarge language modelFinal confidence estimateAgricultural domain knowledge baseIndicator functionDynamically retrieved knowledge for text xInitial judgment result setPrompt template for perspective Arbitration prompt templateAnalysis instruction for perspective Arbitration analysis instructionDescription of the candidate category setNumber of votes for category OOutput format constraintsCategory prediction of perspective Confidence estimate of perspective Global symbol definitions. Methodology Framework overview The multi-perspective prompt fusion framework proposed in this paper aims to enhance the accuracy and robustness of zero-shot agricultural text classification through collaborative judgment from heterogeneous analytical perspectives. The overall architecture of the framework consists of four core processing stages: multi-perspective prompt construction, large language model parallel reasoning and response parsing, consistency measurement and fusion decision, and disagreement detection and arbitration. These stages are sequentially connected to form a complete classification pipeline, where the input text undergoes layer-by-layer processing to output the final category prediction and its confidence estimate. As illustrated in Fig. 1, the multi-perspective prompt construction stage is responsible for generating heterogeneous analysis instructions. For an input text , the framework constructs prompt templates according to M analytical perspectives in the perspective set . Each perspective-specific prompt template contains perspective-specific analysis instructions, description information of the candidate category set , and output format constraints, guiding the large language model to perform classification judgment on the same text from differentiated semantic dimensions. The five perspectives focus on lexical features, semantic associations, sentiment orientation, contextual information, and thematic attribution respectively, forming a complementary analytical perspective system. Overall architecture of the multi-perspective prompt fusion framework for zero-shot agricultural text classification. The large language model parallel reasoning and response parsing stage is responsible for obtaining classification judgments from each perspective. The framework invokes the large language model in a parallel manner, where each perspective prompt is independently input to the model to obtain corresponding natural language responses. The output format constraints in the prompt templates ensure that model responses contain explicit category predictions and self-reported confidence levels. The response parser extracts structured results from each perspective’s output according to preset formats, obtaining the category prediction and confidence estimate for perspective . The consistency measurement and fusion decision stage is responsible for integrating multi-perspective judgments. The framework calculates the perspective consistency metric η based on the prediction results from M perspectives, which quantifies the degree of agreement among perspectives on category attribution judgments. When perspectives exhibit high consistency, the fusion function g employs a majority voting mechanism to determine the final prediction . When disagreement exists among perspectives, the fusion function switches to a confidence-weighted strategy, performing weighted fusion of prediction results by integrating the confidence levels from each perspective. The disagreement detection and arbitration stage is responsible for handling low-confidence samples. The framework compares the consistency metric η with a preset threshold τ. When , the current sample is determined to have significant perspective disagreement. For such samples, the framework triggers a secondary comprehensive judgment mechanism or outputs an uncertainty warning to avoid misleading predictions under perspective conflict situations. This mechanism provides explicit indication of classification result reliability for downstream applications. The above four stages constitute an organic whole. Multi-perspective prompt construction provides heterogeneous information sources for classification, parallel reasoning and response parsing achieve efficient multi-path prediction acquisition, consistency fusion consolidates multi-perspective judgments into unified output, and disagreement arbitration provides quality assurance for the fusion process. The collaborative operation of these modules enables the framework to improve classification performance through perspective complementarity and consistency constraints while maintaining zero-shot characteristics. Multi-perspective prompt design Multi-perspective prompt design aims to construct a differentiated analysis instruction system that guides large language models to perform classification judgment on agricultural texts from multiple semantic dimensions. Single-perspective prompts are insufficient to fully capture the multi-dimensional semantic features of agricultural texts, as different perspectives exhibit varying analytical emphases on the same text. This paper designs five complementary analytical perspectives, each focusing on a specific semantic analysis dimension as in Fig 2. Hierarchical structure of five complementary analytical perspectives, progressing from local to global scope and from explicit to implicit information. The selection of perspectives follows the principle of hierarchical coverage in text analysis. Text understanding typically involves multiple cognitive levels ranging from lexical recognition to discourse comprehension, with information captured at each level possessing certain complementarity and low redundancy. The five-perspective design in this paper corresponds to this hierarchical structure: the lexical perspective corresponds to terminology feature analysis at the lexical level, the semantic perspective corresponds to holistic meaning representation at the semantic level, the sentiment perspective corresponds to attitude and stance judgment at the pragmatic level, the contextual perspective corresponds to disambiguation constraint information at the syntactic level, and the thematic perspective corresponds to macro-level topic attribution at the discourse level. These five levels form complementary coverage from local to global and from explicit to implicit, achieving comprehensive representation of agricultural texts through the synergistic effect of heterogeneous prompts. Accurate classification of agricultural texts depends on the effective utilization of domain knowledge. To enhance the domain adaptability of prompt templates, this paper employs a dynamic knowledge retrieval mechanism based on input text. Let denote the agricultural domain knowledge base, which contains structured knowledge entries including definitions of agricultural terminology, descriptions of crop pest and disease characteristics, and key points of agricultural policies. For an input text x, the framework retrieves knowledge fragments relevant to x from based on semantic similarity. This mechanism enables prompt templates to acquire targeted domain background information, compensating for the deficiencies of large language models in agricultural professional knowledge. The formalized structure of prompt templates is uniformly represented as: where denotes the analysis instruction specific to perspective , denotes the dynamically retrieved external knowledge, denotes the description information of the candidate category set , and O denotes the output format constraints. The five perspectives share three components: , , and O, with only the analysis instruction varying according to different perspectives. The output format constraint O requires the large language model to return category predictions and confidence estimates in a preset format, ensuring the parsability of response results. The design of five analytical perspectives follows the principle of semantic dimension differentiation. The lexical perspective focuses on the domain attribution of specialized terminology in the text, with analysis instruction guiding the model to identify agricultural terms and analyze the indicative strength of term combinations for categories. The semantic perspective concerns the degree of matching between the overall text semantics and the definitions of candidate categories, with analysis instruction guiding the model to evaluate the compatibility between text semantic connotations and category semantic boundaries. The sentiment perspective analyzes the implicit attitudes, stances, and sentiment orientations in the text, with analysis instruction guiding the model to capture potential associations between sentiment features and categories. The contextual perspective focuses on disambiguation information of terminology in specific contexts, with analysis instruction guiding the model to analyze the constraining effect of context on term category orientation. The thematic perspective analyzes the core theme of the text from a macro level, with analysis instruction guiding the model to evaluate the consistency between the text theme and category thematic domains. The above five perspectives constitute a complementary analytical system. The lexical and semantic perspectives conduct analysis from two levels–local terminology features and holistic semantic representation respectively–forming complementarity in analytical granularity. The sentiment perspective captures implicit sentiment information that is difficult to cover through explicit semantic analysis, providing additional discriminative features for classification judgment. The contextual perspective resolves semantic ambiguity of terminology through contextual constraints, improving classification accuracy in polysemous terminology scenarios. The thematic perspective exercises control at the macro theme level, avoiding overall deviation caused by local feature analysis. Compared to the four single-perspective strategies employed by ChatAgri, this paper achieves structured capture of multi-dimensional semantic information through hierarchical perspective design. Compared to Self-Consistency’s reliance on random perturbations introduced by temperature sampling, the perspective differences in this paper originate from different analytical levels, facilitating complementary verification among heterogeneous perspectives. Consistency measurement and fusion strategy After parallel reasoning by the large language model, multi-perspective prompts produce M natural language responses. The response parsing stage is responsible for transforming these unstructured outputs into computable structured representations. The output format constraint O requires the large language model to return category predictions and confidence values in a preset format, and the parser extracts category labels and confidence estimates from each perspective’s response according to this format. The confidence reflects the certainty degree of perspective regarding its own classification judgment, which is autonomously assessed and explicitly output by the large language model based on the analysis process. Consistency measurement aims to quantify the degree of agreement among multi-perspective predictions. This paper adopts a comprehensive measurement scheme combining voting consistency and confidence consistency. Voting consistency measures the concentration degree of category predictions among perspectives, defined as the proportion of perspectives supporting the category receiving the most votes. Let the number of votes for category be , then voting consistency is defined as: This metric has a value range of [1/M, 1]. When all perspectives predict the same category, ; when perspective predictions are uniformly distributed, approaches 1/M. Confidence consistency measures the certainty degree of the perspective group supporting the winning category regarding their own judgments. Let the winning category be , and the set of perspectives predicting be , then confidence consistency is defined as: This metric focuses on the perspective group supporting the final prediction, reflecting the confidence level of these perspectives when making consistent judgments. The composite consistency metric η adopts a product form to fuse the decision level and confidence level dimensions: This design has clear semantic motivation. Voting consistency characterizes the degree of consensus among perspectives on discrete category decisions, while confidence consistency characterizes the certainty degree of the perspective group supporting that decision. The two respectively evaluate the reliability of classification results from the decision level and confidence level. The product form requires no additional hyperparameters and ensures that the composite metric exhibits high values only when both aspects simultaneously achieve high levels–insufficiency in either aspect will directly reduce the composite metric. This property aligns with intuitive expectations for high-confidence classification results. The fusion strategy adopts a combined mechanism separating category decision and confidence quantification. The category prediction is determined by majority voting, selecting the category receiving the most perspective support as the final classification result. When multiple categories have equal vote counts, the one with the highest average confidence among these categories is selected. The final confidence estimate adopts a weighted fusion strategy, integrating confidence information from all perspectives: where is the indicator function, taking a value of 1 when the condition holds and 0 otherwise. This formula calculates the average confidence of perspectives supporting the winning category as the final confidence estimate. The two strategies work collaboratively: the majority voting mechanism ensures the robustness of category decisions, while the confidence fusion mechanism provides quantitative indication of classification result reliability. Disagreement detection and arbitration mechanism The disagreement detection and arbitration mechanism aims to handle samples with significant disagreement among multi-perspective judgments, avoiding misleading predictions under perspective conflict situations. This mechanism identifies low-confidence samples based on the composite consistency metric η and improves classification result reliability through secondary arbitration. Disagreement detection is based on consistency threshold judgment. Threshold setting must balance the characteristic differences of various classification tasks. The larger the number of categories K, the higher the probability of natural disagreement among perspectives, and a fixed threshold would cause excessive triggering of arbitration. To this end, this paper adopts an adaptive threshold setting strategy based on the number of categories: where is the baseline threshold. The motivation for this design is that when the category space expands, the probability of natural disagreement among perspectives increases correspondingly, and automatic threshold reduction can adapt to this change, reducing the need for manual parameter tuning across different datasets. When the composite consistency metric satisfies , the system determines that the current sample has significant perspective disagreement and triggers the arbitration process. The arbitration mechanism adopts a secondary judgment strategy based on comprehensive prompts. When disagreement detection is triggered, the framework constructs an arbitration prompt template for secondary deliberation. The arbitration prompt explicitly injects the initial judgment results from the five perspectives as meta-information. Let the initial judgment result set be , the formalized representation of the arbitration prompt is: where is the arbitration analysis instruction, guiding the large language model to make comprehensive adjudication based on multi-perspective analysis conclusions. This mechanism fully utilizes initial disagreement information, making the arbitration process evidence-based. Compared to simple repeated reasoning or weight adjustment, comprehensive prompt arbitration can integrate multi-perspective judgments at a higher semantic level, possessing stronger disagreement resolution capability. The framework’s final output adopts a tiered confidence labeling system. Based on the processing path of samples in the classification pipeline, output results are divided into three confidence levels. High-confidence samples are those satisfying in the initial fusion, where multi-perspective judgments are highly consistent and classification results have high reliability. Medium-confidence samples are those that did not reach the threshold in initial fusion but obtained effective adjudication through secondary arbitration, where initial disagreement existed but the arbitration mechanism successfully integrated multi-perspective information. Low-confidence samples are those for which secondary arbitration still failed to achieve effective adjudication, which may involve inherent category ambiguity in the text itself or complex situations beyond the current framework’s processing capability. The tiered confidence output provides fine-grained reliability indication of results for downstream applications, enabling users to adopt differentiated processing strategies for different confidence levels, such as manual review of low-confidence samples or marking them as pending categories. Experiments Experimental setup This study selects three public datasets, PestObserver-France, HumAID, and Amazon-Food-Comments, to construct a cross-lingual, cross-domain, and cross-task-type evaluation framework. PestObserver-France originates from French plant health bulletins, annotated by agricultural domain experts. The task involves identifying two categories of plant health risks mentioned in the text: bioagressor and disease. The text content covers agricultural professional information such as crop growth descriptions and pest control recommendations. This dataset is highly aligned with agricultural text classification tasks, and as the evaluation data for the ChatAgri baseline, it ensures fairness in comparison. HumAID was released by the Qatar Computing Research Institute, containing 77,196 manually annotated English tweets collected from 19 real natural disaster events between 2016 and 2019. The annotation scheme covers 11 humanitarian information categories including infrastructure damage, casualty reports, rescue activities, and warning advisories. Natural disasters have direct impacts on agricultural production, and this dataset can verify the applicability of the method in agricultural disaster early warning scenarios. Its larger number of categories K=11 can also examine the effectiveness of the adaptive threshold mechanism . Amazon-Food-Comments was collected by ChatAgri from the Amazon food review platform and covers three sentiment categories of positive, neutral, and negative, with approximately 200,000 English texts in total. The dataset is not directly tied to agricultural production but serves as one of the original ChatAgri evaluation datasets and has been widely used in cross-lingual agricultural text classification studies. Its sentiment classification task can also examine the effectiveness of the proposed method under implicit sentiment judgment scenarios. This study uniformly samples 5,000 instances from the dataset to ensure category balance and to control API call cost. The three-class setting also serves as an intermediate test case for the adaptive threshold mechanism between the binary and 11-class scenarios. Statistical information of the three datasets is presented in Table 2. Table 2Statistical information of experimental datasets.DatasetSamplesCategories KAvg. LengthLanguagePestObserver-France∼20002∼150FrenchHumAID7719611∼25EnglishAmazon-Food-Comments50003∼80English Statistical information of experimental datasets. Baseline methods cover four categories to systematically evaluate the performance advantages of the proposed method. Traditional machine learning methods include SVM and Random Forest with TF-IDF feature representation, establishing the lower bound of classification performance. Pre-trained model fine-tuning methods include BERT and mBERT, representing the performance upper bound of supervised paradigms, used to evaluate the gap between zero-shot methods and supervised methods. Zero-shot prompting methods include Direct Prompting, Chain-of-Thought, and Plan-and-Solve, all employing single-perspective single-path reasoning, used to verify the performance gains of multi-perspective design. Multi-path consistency methods include Self-Consistency and ChatAgri as core comparison baselines. The former improves prediction robustness through random sampling diversity, while the latter is a representative work in agricultural text classification. The number of sampling paths for both is set to 5 to ensure fair comparison with the five-perspective design in this paper. All large language model methods uniformly adopt GPT-3.5-Turbo with temperature parameter set to 0.7. Complete experimental configuration including LLM settings, framework hyperparameters, baseline settings, and evaluation protocol is summarized in Supplementary Table A1. The selection of evaluation metrics follows the principle of combining classification performance, confidence calibration, and prediction diversity. Classification performance employs Accuracy and Weighted-F1, where the former intuitively reflects overall prediction correctness, and the latter comprehensively integrates precision and recall through category sample weighting, effectively addressing the class imbalance problem in the HumAID dataset. Confidence calibration employs Expected Calibration Error (ECE) to measure the deviation between model output confidence and actual accuracy, which corresponds to the tiered confidence output mechanism designed in the methodology section and can verify whether the confidence estimation constructed based on consistency measurement is reliable. Prediction diversity employs pairwise prediction agreement rate to measure the prediction correlation among analysis units in multi-perspective or multi-path methods, defined as the proportion of any two perspectives or paths predicting the same category across all samples, where a lower agreement rate indicates higher prediction diversity. Statistical significance testing employs paired t-test with significance level of 0.05, and all experiments are repeated 5 times with mean and standard deviation reported to evaluate result stability. To control randomness and quantify API variance, each of the five repeated runs uses a distinct random seed from applied to all random sampling components of the framework, while the LLM temperature is fixed at 0.7 across all methods for fair comparison. The combined variance from API output stochasticity and seed-induced fluctuation is jointly captured by the reported standard deviations, which range from to across all methods and datasets. The proposed method exhibits smaller standard deviations ( to ) than single-path zero-shot baselines ( to ), indicating that the consistency-based fusion mechanism averages out individual API call fluctuations. Regarding inference cost, the proposed method averages approximately 5.2 LLM calls per sample–five parallel perspective calls plus one arbitration call triggered on about 20% of samples–which is on the same order as the 5-path Self-Consistency and ChatAgri baselines. Given the 3.2% to 3.7% accuracy improvement over ChatAgri achieved within a comparable call budget, the proposed method maintains a favorable accuracy-cost trade-off relative to multi-path baselines. Experimental results This section reports the experimental results of the proposed method on three public datasets, including six parts: main experiment comparison, perspective ablation analysis, perspective heterogeneity verification, component ablation analysis, parameter sensitivity analysis, and case study. Main experiment results To verify the overall performance advantages of the proposed method, this section compares it with four categories of baseline methods, including traditional machine learning methods, pre-trained fine-tuning methods, zero-shot prompting methods, and multi-path consistency methods. Table 3 presents the classification performance and confidence calibration results of all methods on PestObserver-France, HumAID, and Amazon-Food-Comments datasets, where Acc. denotes accuracy, W-F1 denotes weighted F1 score, ECE denotes expected calibration error, traditional machine learning methods have no ECE metric as they do not output probability distributions, and optimal results are marked in bold. Table 3Classification performance comparison of all methods on three datasets.PestObserver-FRHumAIDAmazon-FCMethodAcc.F1ECEAcc.F1ECEAcc.F1ECETraditional MLSVM–––Random Forest–––Pre-trained Fine-tuningBERT0.0820.0750.078mBERT0.0950.0880.090Zero-shot PromptingDirect Prompting0.1980.2150.205Chain-of-Thought0.1750.1880.182Plan-and-Solve0.1680.1790.174Multi-path ConsistencySelf-Consistency0.1420.1560.148ChatAgri0.1350.1480.140Proposed MethodOursNote: W-F1 denotes weighted F1; ECE is Expected Calibration Error. “–” indicates not applicable. Best results in bold. Classification performance comparison of all methods on three datasets. Note: W-F1 denotes weighted F1; ECE is Expected Calibration Error. “–” indicates not applicable. Best results in bold. Experimental results demonstrate clear performance hierarchies across different method categories. Traditional machine learning methods, limited by the semantic expression capability of TF-IDF features, achieve approximately 70% accuracy on the binary task PestObserver-France, drop to approximately 47% on the 11-class task HumAID, and reach approximately 58% on the three-class sentiment task Amazon-Food-Comments. Supervised fine-tuning methods BERT and mBERT leverage annotated data for task adaptation, achieving 89.2%, 76.8%, and 86.8% accuracy respectively, constituting the performance upper bound reference. Among zero-shot prompting methods, Direct Prompting, Chain-of-Thought, and Plan-and-Solve show progressive improvements, with the best-performing Plan-and-Solve achieving 77.2%, 59.4%, and 74.6%. Multi-path consistency methods Self-Consistency and ChatAgri further improve performance by aggregating multiple reasoning results, with ChatAgri achieving 82.1%, 64.8%, and 78.6%. The proposed method achieves 85.3%, 68.5%, and 82.1% accuracy on the three datasets, improving 3.2%, 3.7%, and 3.5% compared to the strongest baseline ChatAgri, while the ECE metric decreases from 0.135, 0.148, and 0.140 to 0.098, 0.112, and 0.105, corresponding to relative reductions of 27%, 24%, and 25%. A gap of approximately 3.9% to 8.3% remains between the proposed method and supervised fine-tuning methods, consistent with the inherent limitation of the zero-shot paradigm lacking task-specific supervision signals. The visual comparison in Fig. 3 clearly illustrates the performance hierarchy across method categories and demonstrates that the proposed method achieves the best performance among all zero-shot approaches across all three datasets while narrowing the gap with supervised fine-tuning methods. Classification accuracy comparison across different method categories on PestObserver-France, HumAID, and Amazon-Food-Comments datasets. Perspective ablation analysis To verify the independent contribution and complementarity of the five analytical perspectives, this section designs two types of ablation configurations: single-perspective experiments and removal experiments. Single-perspective experiments retain only a specific perspective for classification to evaluate its independent capability, while removal experiments eliminate a specific perspective to evaluate its marginal contribution. Table 4 presents the perspective ablation experiment results. Table 4Perspective ablation experiment results.PestObserver-FranceHumAIDAmazon-Food-CommentsAblation VariantAcc.W-F1Acc.W-F1Acc.W-F1Lexical OnlySemantic OnlySentiment OnlyContextual OnlyThematic Onlyw/o Lexicalw/o Semanticw/o Sentimentw/o Contextualw/o ThematicFull Method Perspective ablation experiment results. Single-perspective experiment results indicate clear differences in the independent classification capability of each perspective, and the relative ranking of perspectives varies across tasks. On PestObserver-France and HumAID, the semantic perspective performs best, achieving 77.8% and 59.4% respectively, as it directly captures core semantic information of the text. The sentiment perspective performs weakest on these two datasets, achieving only 73.6% and 54.2%, because the sentiment polarity discriminability is relatively low in agricultural pest and disaster information classification tasks. On Amazon-Food-Comments, however, the relative ranking is reversed: the sentiment perspective achieves the highest single-perspective accuracy of 75.2%, surpassing the semantic perspective at 73.8%, because sentiment classification is the native task and sentiment-oriented analysis is naturally aligned with the label space. Removal experiment results show consistent patterns aligned with single-perspective rankings. On PestObserver-France and HumAID, removing the semantic perspective causes the largest performance drop, with 2.7% decrease and 3.1% decrease respectively, confirming the central role of the semantic perspective on these tasks. On Amazon-Food-Comments, removing the sentiment perspective causes the largest drop of 2.0%, confirming its central role on the sentiment task. The complete five-perspective configuration improves 7.5% to 9.1% over the best single-perspective configuration on PestObserver and HumAID, and 6.9% over the best single-perspective configuration on Amazon, verifying that the gains from multi-perspective fusion originate from complementarity among perspectives. The cross-dataset comparison further indicates that no single perspective is uniformly optimal across tasks, and the multi-perspective design maintains stable discriminative power across diverse task distributions. Fig. 4 provides comprehensive visualization of perspective analysis, including heterogeneity comparison, pairwise agreement patterns, individual perspective performance, and ablation impact assessment. Perspective analysis results: (a) heterogeneity comparison between our method and Self-Consistency across three datasets, (b) pairwise agreement matrix among five perspectives on PestObserver-France, (c) single perspective performance across three datasets, and (d) performance drop when removing each perspective across three datasets. Perspective heterogeneity verification To verify the advantages of heterogeneous perspective design over random sampling in terms of prediction diversity, this section comparatively analyzes the prediction correlation between the five-perspective method in this paper and the five-path sampling of Self-Consistency. Prediction correlation is measured by pairwise prediction agreement rate, where a lower agreement rate indicates higher diversity among perspective or path predictions, lower correlation in error patterns, and thus richer complementary information for fusion. Table 5 presents the prediction correlation comparison results of both methods on three datasets. Table 5Prediction correlation comparison between heterogeneous perspectives and random sampling.MethodPestObserver-FranceHumAIDAmazon-Food-CommentsAgreementStd.AgreementStd.AgreementStd.Self-Consistency (5 paths)76.8%74.5%78.2%Ours (5 perspectives)58.4%54.2%56.8%Difference-18.4%--20.3%--21.4%- Prediction correlation comparison between heterogeneous perspectives and random sampling. Experimental results show that the prediction agreement rate of the heterogeneous perspective design in this paper is consistently lower than that of Self-Consistency random sampling across all three datasets. On PestObserver-France, the average pairwise agreement rate of Self-Consistency five paths is 76.8%, while that of the proposed five perspectives is 58.4%, a reduction of 18.4 percentage points. On HumAID, the two are 74.5% and 54.2% respectively, with a difference of 20.3 percentage points. On Amazon-Food-Comments, the two are 78.2% and 56.8% respectively, with a difference of 21.4 percentage points. The three datasets consistently exhibit a 18 to 21 percentage point gap, supporting the stable diversity advantage of heterogeneous perspective design over random sampling. The high agreement rate of Self-Consistency indicates that the random perturbations introduced by temperature sampling are mainly reflected at the expression level, while each path still exhibits high correlation in category decisions. This means that when one path produces misjudgment, other paths are likely to make the same error. The heterogeneous perspective design in this paper achieves improvement in prediction diversity through differentiation in analytical levels, with each perspective approaching analysis from five different cognitive levels: lexical, semantic, sentiment, contextual, and thematic, correspondingly reducing the correlation in error patterns. To further analyze the degree of differentiation among perspectives, Table 6 presents the pairwise prediction agreement rate matrix among the five perspectives in this paper, with data calculated based on the PestObserver-France dataset. Table 6Pairwise prediction agreement rate matrix among five perspectives (%).LexicalSemanticSentimentContextualThematicLexical-62.351.258.755.4Semantic62.3-54.865.161.2Sentiment51.254.8-52.653.1Contextual58.765.152.6-59.8Thematic55.461.253.159.8- Pairwise prediction agreement rate matrix among five perspectives (%). The agreement rate matrix presents a clear hierarchical structure. The semantic and contextual perspectives have the highest agreement rate at 65.1%, as both focus on overall semantic information of the text with partial overlap in analytical dimensions. The sentiment perspective has generally the lowest agreement rate with other perspectives on PestObserver-France, averaging only 52.9%, because the sentiment analysis dimension differs from explicit semantic analysis on plant health bulletin texts. The agreement rate between lexical and semantic perspectives is 62.3%, at a medium level, reflecting the partial correlation between local terminology features and holistic semantic representation. On Amazon-Food-Comments, the pairwise agreement matrix exhibits a similar overall hierarchical structure, but the sentiment perspective’s agreement rate with other perspectives rises to an average of 58.6% due to its native alignment with the sentiment classification task. This cross-dataset variation in matrix structure is consistent with the perspective ranking reversal observed in Section 5.2.2, indicating that the relative differentiation among perspectives is task-dependent. The above results indicate that the five-perspective design achieves effective differentiation in analytical dimensions while maintaining certain correlation, and the complementarity among perspectives provides the foundation for fusion gains. Pairwise agreement alone does not fully characterize ensemble diversity. This paper further reports two complementary diversity metrics: oracle accuracy and pairwise Q-statistic on errors. Oracle accuracy is the upper bound of ensemble accuracy, defined as the proportion of samples for which at least one analysis unit (perspective for the proposed method, path for Self-Consistency) produces the correct prediction. The Q-statistic measures error correlation between pairs of analysis units, where values close to 1 indicate highly correlated errors and values close to 0 indicate independent errors. Table 7 reports the results. Table 7Oracle accuracy and error diversity comparison between heterogeneous perspectives and random sampling.MethodPestObserverHumAIDAmazonOracleQ-statOracleQ-statOracleQ-statSelf-Consistency (5 paths)88.2%0.7272.4%0.6885.6%0.75Ours (5 perspectives)93.5%0.4179.8%0.3891.2%0.44 Oracle accuracy and error diversity comparison between heterogeneous perspectives and random sampling. The proposed method shows a substantially higher oracle accuracy than Self-Consistency on all three datasets, with differences of 5.3%, 7.4%, and 5.6% respectively. This indicates that the heterogeneous perspectives collectively cover a broader space of correct predictions than random sampling paths, confirming that the diversity among perspectives is useful rather than merely noisy. The pairwise Q-statistic on errors is also substantially lower for the proposed method (0.38–0.44 versus 0.68–0.75), meaning that when perspectives make mistakes, they tend to make different mistakes. Together, the higher oracle accuracy and lower error correlation provide direct evidence that the diversity of heterogeneous perspectives is effective rather than superficial, complementing the pairwise agreement rate comparison in Table 5. The gap between the proposed method’s actual accuracy (85.3%, 68.5%, 82.1%) and the oracle accuracy (93.5%, 79.8%, 91.2%) also indicates remaining headroom for further improvement of the fusion mechanism, identified as future work in Section 6. Component ablation analysis To verify the independent contribution of each core component of the method, this section conducts ablation experiments from four dimensions: knowledge retrieval, consistency measurement, fusion strategy, and arbitration mechanism. Knowledge retrieval ablation compares three configurations: no knowledge, static template, and dynamic retrieval. Consistency measurement ablation compares voting consistency , confidence consistency , and combined measurement. Fusion strategy ablation compares majority voting, confidence weighting, and combined fusion. Arbitration mechanism ablation compares configurations with and without arbitration. Table 8 presents the component ablation experiment results. Table 8Component ablation experiment results.Ablation VariantPestObserver-FranceHumAIDAmazon-Food-CommentsAcc.W-F1Acc.W-F1Acc.W-F1No Knowledge RetrievalStatic Knowledge TemplateDynamic Knowledge RetrievalVoting Consistency OnlyConfidence Consistency OnlyCombined ConsistencyMajority Voting OnlyConfidence Weighting OnlyCombined Fusionw/o Arbitrationw/Arbitration Component ablation experiment results. Experimental results show that each component contributes positively to the final performance, and the contribution patterns are consistent across the three datasets. Knowledge retrieval ablation shows that dynamic retrieval improves 2.9% to 3.7% compared to no knowledge configuration, and 1.6% to 2.1% compared to static template, verifying the necessity of dynamic domain knowledge adaptation. Consistency measurement ablation shows that voting consistency is slightly better than confidence consistency by approximately 0.6% to 0.7%, and combined use of both provides additional improvement of 1.2% to 1.3%, verifying the complementarity of the two measurement types. Fusion strategy ablation shows that confidence weighting is slightly better than majority voting by approximately 0.3% to 0.4%, and combined fusion provides additional improvement of 1.1% to 1.4%. The arbitration mechanism brings 0.8% to 0.9% performance improvement, consistent with its design logic of only taking effect on low-consistency samples. The stable contribution patterns across three datasets with different category granularities and task types indicate that the component design is not over-fitted to any specific task distribution. Parameter sensitivity analysis To analyze the impact of key hyperparameters on method performance and provide parameter selection guidance for practical applications, this section conducts sensitivity experiments from three dimensions: consistency threshold , number of perspectives, and consistency weight α. Table 9 presents the impact of consistency threshold on classification performance and arbitration triggering ratio, with testing range covering 0.3 to 0.8, where Arb. denotes arbitration triggering ratio. Table 9Impact of consistency threshold on performance and arbitration ratio.PestObserver-FranceHumAIDAmazon-Food-CommentsAcc.Arb.Acc.Arb.Acc.Arb.0.384.228.5%66.835.2%80.432.8%0.484.722.3%67.528.6%81.325.1%0.585.316.8%68.521.4%82.118.5%0.685.111.2%68.214.8%81.812.6%0.784.66.5%67.48.9%80.97.4%0.883.92.8%66.54.2%79.83.5% Impact of consistency threshold on performance and arbitration ratio. Experimental results show an inverted U-shaped relationship between threshold setting and performance. A threshold too low causes too many samples to trigger arbitration, introducing additional noise, while a threshold too high causes truly difficult samples requiring arbitration to go unidentified. The optimal threshold achieves best performance on all three datasets, with arbitration triggering ratios of 16.8%, 21.4%, and 18.5% respectively, all maintained in the reasonable range of 16% to 22%. The arbitration triggering ratio scales monotonically with the number of categories: 16.8% for the binary task PestObserver, 18.5% for the three-class task Amazon, and 21.4% for the 11-class task HumAID, validating the design intuition that perspective disagreement is more likely in larger label spaces and that the adaptive threshold mechanism correctly accommodates this scaling. Figure 5 provides comprehensive visualization of component ablation and parameter sensitivity analysis, clearly demonstrating the contribution of each component and the optimal parameter configurations. Component ablation and parameter sensitivity analysis: (a) component ablation results across three datasets, (b) threshold sensitivity with accuracy and arbitration rate, (c) performance variation with number of perspectives, and (d) weight sensitivity between voting and confidence consistency. Table 10 presents the impact of number of perspectives on classification performance, revealing the law of diminishing marginal returns in performance gains by progressively increasing the number of perspectives. Table 10Impact of number of perspectives on classification performance.Number of PerspectivesPestObserver-FranceHumAIDAmazon-Food-Comments176.257.873.0280.562.476.8383.165.679.5484.567.481.2585.368.582.1 Impact of number of perspectives on classification performance. Experimental results present a clear trend of diminishing marginal returns. The increment from 1 to 2 perspectives is 3.8% to 4.6%, from 2 to 3 perspectives is 2.6% to 3.2%, from 3 to 4 perspectives is 1.4% to 1.8%, and from 4 to 5 perspectives decreases to 0.8% to 1.1%. Single-perspective configuration performance is approximately 76.2%, 57.8%, and 73.0% on the three datasets, close to Direct Prompting, verifying the necessity of multi-perspective design. The 5-perspective configuration achieves a reasonable balance between performance and computational cost. Table 11 presents the impact of consistency weight α on classification performance, where the composite consistency metric is generalized as , α =0 indicates relying solely on confidence consistency, α =1 indicates relying solely on voting consistency, and α =0.5 reduces to the balanced product form adopted in the methodology. Table 11Impact of consistency weight α on classification performance.αPestObserver-FranceHumAIDAmazon-Food-Comments0.083.566.580.30.284.267.380.80.484.868.181.60.585.368.582.10.685.168.381.90.884.467.681.21.084.167.280.7 Impact of consistency weight α on classification performance. Experimental results show that equal-weight combination of the two consistency measurements is an effective strategy across all three datasets. Extreme weights α =0 and α =1 both cause performance degradation, verifying the necessity of complementarity between the two measurements. The performance curve shows approximately symmetric distribution centered at α =0.5 on all datasets. The optimal weight α =0.5 indicates that voting consistency and confidence consistency contribute comparably, allowing direct application to different task scenarios without complex weight tuning. Case study To illustrate how the multi-perspective prompt fusion framework operates on a text with mixed surface cues, this section presents a representative case from the Amazon-Food-Comments dataset where single-perspective prompting tends to misclassify the sample but the proposed framework produces the correct prediction through perspective complementarity. The input review reads: “One of my boys needed to lose some weight and the other didn’t. I put this food on the floor for the chubby guy.” The ground-truth label is positive. The reviewer does not use explicit positive sentiment words such as “love” or “great”; instead, positive evaluation is implicit through the successful use of the product for a weight management purpose. Superficially, the text contains lexical items with negative associations such as “lose weight” and “chubby” that can mislead sentiment-oriented analysis. Table 12 reports the five-perspective outputs, the consistency metrics, and the final decision. Table 12Case study: multi-perspective analysis of an Amazon food review with implicit positive sentiment.PerspectivePredictionConfidence Lexicalneutral0.55Semanticpositive0.72Sentimentnegative0.58Contextualpositive0.75Thematicpositive0.68Voting consistency 0.60Confidence consistency 0.72Composite consistency η0.432Adaptive threshold 0.289Arbitration triggeredNoInitial fused predictionpositiveGround truthpositive Case study: multi-perspective analysis of an Amazon food review with implicit positive sentiment. The five perspectives split three-to-two in favor of the positive class. The lexical perspective returns neutral because the surface vocabulary contains no explicit polarity-bearing tokens. The sentiment perspective returns negative because it is misled by the surface occurrences of “lose weight” and “chubby”, treating them as dispreferred states. The semantic, contextual, and thematic perspectives correctly infer the implicit positive evaluation. The semantic perspective captures the purposive use of the product for weight management. The contextual perspective recognizes that “chubby guy” refers to the target user of the product rather than a complaint. The thematic perspective identifies the overall theme as a product-use report with implicit endorsement. Notably, the sentiment perspective, which is the strongest performer in isolation on this dataset as shown in Section 5.2.2, produces an incorrect judgment on this particular sample, illustrating that even the task-aligned perspective benefits from cross-validation by other perspectives. Under majority voting, the fused prediction is positive with fused confidence 0.72, aligned with the ground truth. The composite consistency η = 0.432 exceeds the adaptive threshold τ = 0.289 for the three-class task, so no arbitration is triggered, and the sample is placed in the high-confidence tier. This case illustrates three design points. First, heterogeneous perspectives produce genuinely divergent judgments on a text where surface cues diverge from the underlying evaluative stance, and the lexical and sentiment perspectives fail in different ways that cannot be corrected by duplicating similar perspectives through sampling. Second, the semantic, contextual, and thematic perspectives provide complementary evidence at different levels of abstraction, converging on the correct label despite each operating on a different aspect of the text. Third, the composite consistency correctly reflects that two perspectives dissented, yielding a moderate rather than saturated confidence value, which is consistent with the actual difficulty of the sample. A second representative case that triggers arbitration, together with an additional failure case where the framework produces an incorrect prediction, is provided in the Supplementary Information (Supplementary Tables B1 and B2) for completeness. Discussion The experimental results reveal the performance boundaries of different technical approaches in zero-shot agricultural text classification tasks. Traditional machine learning methods can achieve approximately 70% accuracy on binary classification tasks, but performance drops sharply to around 47% when facing 11-class fine-grained classification, exposing the limitations of shallow feature representations. Supervised fine-tuning methods occupy the performance upper bound due to their advantage of adapting with annotated data, but the scarcity of annotation resources in the agricultural domain limits their practical application. Zero-shot prompting methods show steady improvement along the evolutionary trajectory from Direct Prompting to Plan-and-Solve. Self-Consistency and ChatAgri further break through the single-path bottleneck through multi-path aggregation, while the additional 3.2% to 3.7% improvement achieved by the proposed method indicates that perspective heterogeneity has certain advantages over sampling randomness. The effectiveness of the multi-perspective fusion strategy is rooted in the core principle of base learner diversity in ensemble learning. Self-Consistency relies on temperature sampling to generate multiple reasoning paths, but the diversity under the constraint of the same prompt template only originates from random noise at the decoding level, and different paths tend to make the same errors on similar samples. The perspective heterogeneity verification experiment shows that the average pairwise prediction agreement rate of Self-Consistency’s five paths is approximately 75%, while that of the five perspectives in this paper is approximately 56%, with a difference of 18 to 21 percentage points. The proposed method constructs a heterogeneous analysis framework based on the principle of hierarchical coverage in text analysis, where each perspective approaches analysis from different cognitive levels, error patterns exhibit lower correlation, and misjudgments from one perspective are more likely to be corrected by other perspectives. In the perspective ablation experiment, the complete configuration achieves 7.5% to 9.1% improvement compared to the best single perspective, and removing any perspective leads to performance degradation, providing direct evidence for perspective complementarity. The effectiveness of consistency measurement is established on a straightforward observation: when multiple independent sources reach a high degree of consensus, the reliability of that conclusion is typically higher than any single source alone. Voting consistency characterizes the degree of consensus among perspectives at the decision level, while confidence consistency characterizes the certainty degree of predictions at the confidence level. The two evaluate the reliability of classification results from different dimensions respectively, and combined use can construct a more complete reliability assessment. Component ablation experiments show that the combined measurement provides an additional 1.2% to 1.3% improvement compared to single measurements, verifying the complementarity of the dual-dimensional design. The contribution of dynamic knowledge retrieval originates from targeted compensation for the shortcomings of LLMs in agricultural professional knowledge, customizing targeted knowledge for each input through semantic similarity matching, making the injected information highly relevant to the current task. The arbitration mechanism identifies low-consistency difficult samples and makes prudent judgments by integrating reasoning evidence from each perspective. Approximately one-fifth of samples trigger arbitration and bring nearly 1% performance gain, indicating that this mechanism precisely handles the boundary sample subset. The practical value of the proposed method is manifested in typical pain point scenarios of agricultural informatization. The zero-shot characteristic enables deployment without task-specific annotation, significantly lowering the application barrier. Confidence estimation supports differentiated processing strategies for high and low confidence samples. The multi-perspective framework provides natural interpretability, with outputs and reasoning processes from each perspective being traceable. Adaptation to new tasks can be achieved by adjusting category labels and knowledge bases without retraining, and this flexibility is particularly valuable in rapidly changing application environments. The framework also has several limitations that point to directions for future work. First, the gap between the actual fused accuracy and the oracle accuracy reported in Table 7 indicates that the current fusion mechanism does not fully exploit the complementarity among perspectives, leaving room for more expressive fusion designs beyond the current consistency-based voting. Second, as illustrated in the failure case in the Supplementary Information, the consistency-based reliability indicator cannot detect samples where all perspectives share the same systematic bias and converge on the same incorrect prediction, suggesting the need for an additional perspective weighted by aspect importance or an external calibration mechanism that leverages sample-level features. Third, evaluation with a dedicated agricultural-domain large language model is currently constrained by the lack of a publicly released agricultural LLM of sufficient scale, and this evaluation is left for when such a model becomes available. The experimental verification covers three datasets across different languages, category granularities, and task types. Evaluation on larger-scale benchmarks and additional agricultural-specific corpora is left for future work. Conclusion This paper addresses the problem that single-perspective prompts in zero-shot agricultural text classification tasks cannot fully cover multiple levels of text analysis and lack cross-validation mechanisms, proposing a multi-perspective prompt fusion framework. This framework constructs five complementary analytical perspectives–lexical, semantic, sentiment, contextual, and thematic–based on hierarchical differences in text analysis, forming multi-level coverage from local terminology features to overall thematic attribution. It guides large language models to perform classification judgment from different cognitive dimensions, constructs a dual-dimensional consistency measurement mechanism integrating decision-level and confidence-level metrics to achieve reliability assessment of multi-perspective judgments, and prudently handles low-consistency samples through adaptive threshold-driven disagreement detection and arbitration mechanisms. Experiments on the PestObserver-France, HumAID, and Amazon-Food-Comments datasets demonstrate that the proposed method achieves 3.2% to 3.7% accuracy improvement and 24% to 27% reduction in confidence calibration error compared to ChatAgri. Ablation experiments verify the positive contribution of each component. Perspective heterogeneity verification shows that the prediction agreement rate of the proposed method is lower than that of random sampling methods by 18 to 21 percentage points, with oracle accuracy exceeding Self-Consistency by 5.3% to 7.4% and pairwise Q-statistic on errors reduced from 0.68–0.75 to 0.38–0.44, providing quantitative support for the complementarity among perspectives. Experimental results indicate that perspective heterogeneity has advantages over sampling randomness in multi-path fusion, providing an effective solution for zero-shot text classification in the agricultural domain without requiring domain-specific annotation. Supplementary Information Supplementary Information. Author contributions Q.L. and Z.X. conceived the research idea and designed the multi-perspective prompt fusion framework. Q.L. implemented the methodology, conducted the experiments, and performed the data analysis. Y.D. curated the datasets, carried out text preprocessing, and assisted with experimental validation. H.L. contributed to the literature review and provided critical interpretation of results. Z.X. formulated the theoretical framework, supervised the technical development, and revised the manuscript. Y.Z. supervised the overall research, and provided strategic guidance. Q.L. and Z.X. wrote the main manuscript text. Y.D. prepared Figs. 1–3 and Tables 1–4. H.L. and Y.Z. substantively revised the manuscript for intellectual content. All authors reviewed and approved the final manuscript. Funding This work was supported by the Guangxi Key Laboratory of Machine Vision and Intelligent Control (Grant No. 2025A02) Data availability The PestObserver-France dataset is available from the ChatAgri study (Zhao et al., 2023). The HumAID dataset is publicly available at https://crisisnlp.qcri.org/humaid_dataset. The experimental results generated during this study are included in this published article. Declarations Competing interests The authors declare no competing interests. Ethical approval This study is purely computational and does not involve human participants, human tissue samples, or personal data. Ethical approval is therefore not applicable. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Yaoyao Ding, Email: dingyaoyao@cuz.edu.cn. Zeling Xu, Email: ZelingXu@ieee.org. Yuntao Zou, Email: zouyuntao@hust.edu.cn. Supplementary Information The online version contains supplementary material available at https://doi.org/10.1038/s41598-026-56518-1. References Associated Data Supplementary Materials Supplementary Information. Data Availability Statement The PestObserver-France dataset is available from the ChatAgri study (Zhao et al., 2023). The HumAID dataset is publicly available at https://crisisnlp.qcri.org/humaid_dataset. The experimental results generated during this study are included in this published article.