Rice disease and insect pest question-answering method based on circular retrieval

Abstract

The invention relates to the field of rice disease and insect pest questioning and answering, in particular to a rice disease and insect pest questioning and answering method based on circular retrieval. The scheme comprises the following steps: collecting data; constructing a rice disease and insect pest knowledge base; encoding a rewritten problem into vector representation by using an embedded model; executing cosine similarity matching in the knowledge base; whether the retrieved context can answer the question of the user or not is judged, if yes, the answer to the question is output through a large language model, if not, the large language model is used as a generation model, a targeted reverse question is generated in combination with prompt words, and the user is guided to provide more rice disease information, environmental factors and symptom description in a reverse question mode; and then integrating and optimizing to obtain a new question, executing cosine similarity matching in the knowledge base through the obtained question, and realizing efficient and accurate answering through the circular retrieval. The method is suitable for rice disease and insect pest questioning and answering.

Description

Technical Field

[0001] The present invention relates to the field of rice pest and disease Q&A, and specifically relates to a rice pest and disease Q&A method based on cyclic retrieval. Background Art

[0002] In the research of rice pest and disease Q&A systems, most are still based on traditional knowledge graph methods. However, the Q&A systems based on knowledge graphs have the following problems: Limited knowledge coverage and insufficient recognition of rare pests and diseases: The knowledge graph depends on the coverage of structured data, and its accuracy and integrity are limited by the pre-built knowledge base. For rare or newly emerging pest and disease types, the knowledge graph may not be able to effectively identify them due to insufficient data samples; Lag in dynamic update and high maintenance cost: The update of the knowledge graph requires manual intervention (such as expert review, data cleaning), and it is difficult to integrate the latest pest and disease research results or field emergencies in real time; Insufficient natural language interaction fluency and difficulty in handling complex contexts: Systems based on knowledge graphs usually rely on keyword matching or limited semantic parsing, and users need to ask questions in a specific format.

[0003] Other traditional RAG (Retrieval-Augmented Generation) models are used to make up for the deficiencies of existing Q&A systems based on large language models in terms of information accuracy and knowledge depth. However, there are still the following problems: Facing users' vague or complex query requests, traditional RAG models passively rely on single retrieval results to generate answers, unable to actively guide users to explore real needs, which easily leads to incomplete retrieval or misinterpretation of user intentions; Traditional RAG models are based on static retrieval and are difficult to handle complex or multi-round scenarios, which may lead to fragmented information or broken context; The interactivity between the generation module and the retrieval module is poor and lacks linkage. Summary of the Invention

[0004] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a rice pest and disease Q&A method based on cyclic retrieval, which greatly improves the accuracy and comprehensiveness of answers.

[0005] The present invention adopts the following technical solutions to achieve the above purpose. The present invention provides a rice pest and disease Q&A method based on cyclic retrieval, including: Step 1, collect text data related to rice; Step 2, perform preprocessing of data cleaning and data structuring on the collected rice data; Step 3, perform chunking and segmentation processing on the preprocessed rice text data; Step 4: Use the embedding model to encode the text data processed in Step 3 into vector representations, and store the vectors in a vector database to construct a vectorized knowledge base for rice pests and diseases; Step 5: After receiving a user's question about rice pests and diseases, first apply the query preprocessing and enhancement function Trewrite to transform it and generate an optimized query; Step 6: Use the embedding model to convert the optimized query into a query vector, and then perform cosine similarity matching in the vectorized knowledge base for rice pests and diseases to find the context related to the user's question; Step 7: Calculate the sufficiency score of the context. This sufficiency score consists of two parts: keyword coverage and content support. Keyword coverage is used to evaluate whether the retrieved content covers the corresponding concepts in the question, and content support evaluates whether the retrieved context can provide direct content basis for answering the user's question through the ROUGE-L metric; Step 8: Compare the sufficiency score with a pre-set threshold, and at the same time make a judgment in combination with the current interaction iteration count and the set maximum iteration rounds. If the sufficiency score is greater than or equal to the set threshold, it is determined that the currently retrieved context information is sufficient, and go to Step 11. If the sufficiency score is less than the set threshold and the current interaction iteration count is less than the set maximum iteration rounds, it is determined that the currently retrieved context information is insufficient, and go to Step 9. If the sufficiency score is less than the set threshold and the current interaction iteration count is greater than or equal to the set maximum iteration rounds, then give the user the set of contexts with the highest score retrieved currently, and prompt the user that the information in the current knowledge base is incomplete or insufficient to form a completely definite suggestion, reminding the user to refer to the provided information with caution; Step 9: Use a large language model as the generation model, and combine with prompt words to generate targeted clarification questions. The prompt words include the current query, the retrieved set of contexts, and the calculated sufficiency score; Step 10: Present the clarification question to the user and receive the user's answer, and then fuse the answer information provided by the user with the current query to generate the next round of optimized query, and return to Step 6; Step 11: Integrate the finally confirmed query and the finally retrieved and confirmed set of contexts to form a complete generation prompt, and input this generation prompt into another large language model to generate the final answer to the user's question about rice pests and diseases.

[0006] Furthermore, step 1 specifically includes: collecting and organizing unstructured text data related to rice pests and diseases from various channels, including scientific research papers on rice pathology, entomology, and pesticides published publicly, technical reports and pest and disease forecasts released by national or local plant protection stations, as well as control guides and variety resistance information released by professional agricultural research institutions.

[0007] Furthermore, step 3 specifically includes: performing block processing on the text data of rice pests and diseases. For the chapters on a certain disease or pest in scientific research papers and technical reports, use the title as the natural segmentation point for blocking. At the same time, construct a white list of professional terms in the field of rice pests and diseases, and refer to this white list during blocking to ensure that text fragments containing these terms are not incorrectly segmented.

[0008] Furthermore, the calculation methods of sufficiency score, keyword coverage rate, and content support degree are as follows: ; ; ; In the formula, represents the keyword coverage rate, represents the content support degree, represents the sufficiency score of the final context, represents and the length of the longest common subsequence, represents the length of represents the retrieved context, represents the optimized query keyword set, represents the keyword set of the retrieved context, and are weight parameters.

[0009] Furthermore, the method for generating the optimized query is as follows: , where represents the optimized query, Q represents the original question input by the user, and Trewrite represents the enhancement function.

[0010] The beneficial effects of the present invention are: The present invention improves the accuracy and coverage of the question-answering system: Traditional knowledge graph question-answering systems are limited by the scope of predefined structured data and it is difficult to cover all pests and diseases, especially rare or newly emerging types. Traditional RAG models directly generate answers after a single retrieval, with insufficient depth of information acquisition. This solution cleverly integrates the powerful reasoning ability of large language models with a multi-round dynamic retrieval mechanism: When the initially retrieved information is insufficient, the system can actively analyze the missing links and guide the user to supplement key information through generative questions. This iterative information collection and reasoning process significantly improves the diagnostic accuracy for complex and ambiguous diseases and effectively expands the system's recognition and coverage capabilities for rare or atypical rice pests and diseases.

[0011] The present invention enhances the user interaction experience and supports the resolution of complex and cross-jump questions: Faced with vague descriptions from users or complex queries involving multiple logical relationships, traditional systems are often at a loss. The core advantage of this solution lies in its interactive clarification and iterative optimization mechanism. The system can dynamically generate highly targeted follow-up questions based on the current retrieval results, actively guiding the user to clarify their true intentions and provide key details. Through this cycle of "judgment - generating questions - iterative retrieval", the context information required to solve the problem is gradually constructed and improved, greatly enhancing the system's understanding depth of complex contexts and its effective response ability to multi-hop reasoning questions, making the user experience closer to a natural conversation with domain experts.

[0012] The present invention improves the adaptability of question-answering and reduces the maintenance cost: The construction and maintenance of knowledge graphs rely on a large amount of expert annotation and structured data processing, which is not only costly but also the knowledge update often lags behind the latest research progress and field actual situations. This solution is based on RAG and naturally has the ability to integrate unstructured text data. This means that the system can more conveniently and quickly access and utilize external data sources such as the latest scientific research literature, plant protection reports, and online information, continuously learning and absorbing new knowledge. This not only significantly reduces the manual maintenance cost and the delay of knowledge update, but also endows the system with excellent adaptability, enabling it to flexibly respond to the rapid development and changes in the knowledge of the rice pests and diseases field. Brief Description of the Drawings

[0013] Figure 1 It is a flowchart of a method for answering questions about rice pests and diseases based on iterative retrieval provided by an embodiment of the present invention. Detailed Embodiments

[0014] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0015] The present invention provides a method for answering questions about rice pests and diseases based on iterative retrieval, as Figure 1As shown below, specifically including: Step 1: Collect and organize unstructured rice text data from scientific research papers, technical reports, and professional agricultural research institutions. Specifically, systematically collect and organize unstructured text data related to rice pests and diseases from multiple authoritative sources. These sources mainly include scientific research papers on rice pathology, entomology, and pesticide science published publicly, technical reports and pest and disease forecasts released by national or local plant protection stations, as well as control guidelines and variety resistance information released by professional agricultural research institutions (such as the China National Rice Research Institute). The goal is to construct a basic dataset that widely and deeply covers knowledge of major rice diseases (such as rice blast, sheath blight, bacterial blight, false smut, etc.) and pests (such as rice planthoppers, striped stem borers, rice leaf folders, etc.).

[0016] Step 2: Conduct preprocessing of data cleaning and data structuring on the collected rice data. Specifically, conduct comprehensive data cleaning and structured preprocessing on the collected original rice pest and disease data. This step aims to eliminate noise in the data, such as format errors, irrelevant information, and duplicate descriptions, etc. And transform unstructured text data, such as disease symptom descriptions, morphological characteristics, occurrence patterns, and recommended control agents, etc. into a format that is more conducive to machine processing, laying a solid foundation for subsequent text analysis and vectorization, and ensuring the professionalism and accuracy of the input data.

[0017] Step 3: Perform chunking on the rice text data. For scientific research papers and technical reports, use title segmentation, and at the same time construct a whitelist of professional terms to retain the complete chapter logical chain and professional terms. Perform chunking on the preprocessed rice pest and disease text data. The specific strategy is to use the title as the natural segmentation point for chunking the chapters in scientific research papers and technical reports regarding a specific disease or pest. At the same time, in order to prevent the integrity of professional terms such as key pest and disease names, active ingredients of pesticides, growth stages, symptom descriptions, etc. from being damaged during the chunking process, a whitelist of professional terms in the field of rice pests and diseases will be specifically constructed. Refer to this whitelist during chunking to ensure that text fragments containing these terms are not incorrectly segmented, thereby maintaining the integrity of the knowledge logical chain and professional terms.

[0018] Step 4: Use the text-embedding-ada-002 embedding model to encode and transform the text data chunks processed in the previous step into high-dimensional vector representations. These vectors can capture the deep semantic information of the text. Subsequently, store these generated text vectors in a vector database to construct a vectorized knowledge base of rice pests and diseases that can be quickly retrieved. Step 5: After receiving the user's question about rice pests and diseases, first apply the query preprocessing and enhancement function Trewrite to transform it, generate an optimized query, standardize the colloquial description, supplement key information (such as the growth stage of rice), or perform synonym expansion of pest and disease names (such as associating "rice blast" with "rice fever").

[0019] The generation formula of the optimized query is as follows: ; Where represents the optimized query, Q represents the original question input by the user; Trewrite represents the function that performs the transformation. Its core goal is to intelligently optimize and standardize the user's original, possibly ambiguous or colloquial query, generating a clearer, more structured and information-rich query version, aiming to improve the efficiency and accuracy of subsequent steps. The specific operations cover multiple aspects: converting non-standard colloquial descriptions into professional agricultural terms, correcting spelling mistakes; expanding keywords using built-in thesaurus or domain knowledge to ensure that different expressions such as "rice blast" and "rice fever" can match the same content; attempting to identify and supplement missing key context information in the query, such as inferring or supplementing the growth stage of the current rice based on the conversation history; and identifying and linking entities such as pests, diseases, pesticides, etc. mentioned in the query to the standard identifiers in the knowledge base. Through this series of transformations, the optimized query can more accurately reflect the user's intention and use language closer to the literature in the knowledge base, thus significantly improving the recall rate and accuracy of subsequent vector retrieval.

[0020] Step 6: Use the specified embedding model to convert the processed optimized query into its high-dimensional representation in the vector space, that is, the query vector . Immediately afterwards, perform an efficient similarity search in the pre-constructed rice pest and disease vector knowledge base. By calculating the cosine similarity between the query vector and each stored text fragment vector in the knowledge base, the system retrieves the top-k context text fragments that are semantically closest to the user's question, forming a preliminary candidate result set . These retrieved fragments are expected to contain core information directly related to the user's query, such as specific pest and disease descriptions, key diagnostic points, or effective control methods.

[0021] Among them, the calculation formula of the cosine similarity is: In the formula, represents the query vector and the i-th stored text fragment vector in the knowledge base The cosine similarity between

[0022] Step 7: To further improve the accuracy of the retrieved context, after obtaining the candidate set B through the preliminary vector retrieval, a re-ranking model is introduced and applied for secondary precise ranking to obtain the retrieved context set . This step aims to make up for the deficiency in the ranking accuracy of the preliminary retrieval and solve the problem that the most relevant documents may not be at the top of the list. The core task of re-ranking is to conduct a deep relevance assessment of the candidate documents through more complex calculations and models, and accurately place the documents that best match the user's query intention at the forefront of the sorted list, thereby significantly improving the precision rate of the top results. In this solution, this precise ranking process is implemented using the Cross-Encoder model: this model can simultaneously receive the optimized query and each candidate document as inputs, carefully analyze the deep semantic interaction relationship between the two, and generate a score accordingly. Finally, the candidate documents are re-ranked from high to low according to this score. After the re-ranking process, the top of the output context list is concentrated with the highest-quality and most directly relevant information, obtaining the retrieved context set , which provides a more solid foundation for the subsequent context sufficiency assessment and final answer generation, effectively guarantees the accuracy of the question and answer, and may effectively reduce unnecessary interaction clarification rounds.

[0023] Step 8: After retrieving the preliminary context set , the core task of this step is to quantitatively evaluate whether this batch of information is sufficient - that is, whether the user's original question can be answered effectively and accurately based on them. For this purpose, the present invention designs a comprehensive sufficiency score, which consists of two key dimensions: Keyword coverage rate: This indicator aims to measure the extent to which the retrieved content covers the core concepts mentioned in the user's question. It ensures the relevance of the retrieval results to the user's concerns and prevents the omission of important information. The specific calculation method is to compare the coincidence degree of the keyword set of the rewritten question and the keyword set of the context. The calculation formula is: ; where, represents the keyword set of the question, represents the keyword set of the retrieved context, represents the keyword coverage rate of the former two.

[0024] Content Support: The design of this metric goes beyond simple relevance judgment and aims to more deeply evaluate whether the retrieved context can provide a direct and substantial content basis for answering the user's question. The present invention selects the method based on the longest common subsequence to quantify this degree of support. Obtaining a higher value generally means that there is a stronger direct connection between the retrieved context and the expression of the user's question, and it is more likely to contain the specific information or key details necessary to generate an accurate and reliable answer.

[0025] Its calculation formula is: ; Where represents and the length of the longest common subsequence, represents the length of represents the content support.

[0026] By fusing these two dimensions with weights, the final context sufficiency score is obtained, which combines the information relevance and the answer generability: ; Wherein, is the sufficiency score between the user's question and the retrieved context, and are weight parameters.

[0027] The weight parameters and reflect the judgment of the respective importance of coverage and support. According to experiments, when α = 0.6 and β = 0.4 are set, it can better balance these two aspects and achieve the best overall evaluation effect. The score will be used as the key basis for the next decision, determining whether to directly generate an answer or further interaction and clarification are needed.

[0028] Step 9. After calculating the final context sufficiency score in the previous step, this step compares this score with a pre-set threshold τ to decide the subsequent action path. At the same time, it makes a judgment in combination with the current interaction iteration number n and the set maximum iteration number for judgment.

[0029] Context Sufficient: If , the system determines that the currently retrieved context is sufficient enough to directly generate a high-quality answer. The process will jump to Step 12.

[0030] Context Insufficient and Can Continue Iterating: If and , indicating that the current context information is still insufficient, but there is still an opportunity to improve through interactive clarification. The process will enter step 10.

[0031] Reached the iteration limit and the context is still insufficient: If and , this means that after the maximum allowed number of rounds of interactive clarification, the retrieved context still fails to meet the preset sufficiency standard. In this specific case, the system will no longer attempt to generate a comprehensive answer. The following special processing will be performed: Summarize the currently retrieved set of contexts with the highest scores to the user. At the same time, the system will clearly inform the user that although it has tried its best to retrieve and clarify, the information in the current knowledge base may still be incomplete or insufficient to form a completely definitive diagnosis or recommendation, prompting the user to refer to the provided information with caution.

[0032] Step 10. When step 9 determines that the currently retrieved context is insufficient to support a high-quality answer and the number of interaction rounds allows for continuation, the system will initiate a clarification question mechanism. This step utilizes Deepseek-7B, combined with the designed prompt engineering. The prompt will include the current query, the retrieved set of contexts, and the calculated sufficiency score . Based on these inputs, the large language model will generate a highly targeted clarification question . This question will focus on guiding the user to provide specific details crucial for the diagnosis or control of rice pests and diseases, such as: "What are the specific shapes, colors, and sizes of the spots on the rice leaves you observed? Is there a yellow halo around the lesions?" or "What is the field humidity in the disease-affected area? Has it been continuously rainy recently?" Step 11. The system presents the clarification question generated in step 10 to the user and receives the user's answer , such as, "The lesions are small round brown spots with a yellow halo around them" or "The field humidity is very high and it has been raining for the past week." After obtaining the feedback, the system applies a specific integration function T to effectively integrate the new, specific pest and disease-related information provided by the user with the current query where is the optimized query generated in this round, T is the integration function that performs information integration, is the new relevant information provided by the user in this round, is the previous query currently being processed.

[0033] Subsequently, increment the interaction iteration counter n by 1 and carry this new query optimized by the user's feedback The entire process will return to step 6 and start executing vectorization, similarity retrieval, re-ranking, and subsequent context evaluation again.

[0034] Step 12: When the judgment condition in step 9 is satisfied, the process enters the stage of generating the final answer. The system will integrate the finally confirmed query and the finally retrieved and confirmed set of contexts used to construct a complete generation prompt. This prompt will be input into another large language model, which will comprehensively reason and generate a final answer to the user's rice pest and disease problem based on the provided query and context information. It should be particularly noted that if the process enters this step because the maximum number of iterations has been reached, but the context sufficiency score at this time is still lower than the threshold, the system should clearly inform the user in the generated answer that the current information based on the knowledge base may not be sufficient to provide a completely conclusive answer, and may suggest that the user seek the help of professional agricultural technicians or provide other auxiliary information. At the same time, the system should record such problems that cannot be completely solved for subsequent update of the knowledge base and model optimization.

[0035] The above are only the preferred embodiments of the present invention. It should be understood that the present invention is not limited to the form disclosed herein, should not be regarded as excluding other embodiments, and can be used in various other combinations, modifications, and environments, and can be changed within the scope of the concept described herein through the above teachings or the technology or knowledge in related fields. Any changes and variations made by those skilled in the art without departing from the spirit and scope of the present invention shall fall within the protection scope of the appended claims of the present invention.

Claims

1. A rice pest question-answering method based on circular retrieval, characterized in that: include: Step 1: Collect text data related to rice; Step 2: Preprocess the collected rice data by cleaning and structuring the data; Step 3, dividing and segmenting the pre-processed rice text data; Step 4: Encode the text data processed in step 3 into vector representation using the embedding model, and store the vector into a vector database to construct a vectorized knowledge base of rice pests and diseases; Step 5: After receiving the user's question about rice pests and diseases, query preprocessing and enhancement functions are applied to transform the query to generate an optimized query; Step 6: Use the embedding model to convert the optimized query into a query vector, and then perform cosine similarity matching in the rice pest and disease vectorized knowledge base to find the context related to the user's question; Step 7: Calculate the context sufficiency score, which is composed of two parts: keyword coverage and content support. Keyword coverage is used to evaluate whether the retrieved content covers the corresponding concepts in the question. Content support is evaluated through the ROUGE-L indicator to evaluate whether the retrieved context can provide direct content basis for answering the user's question. Step 8: Compare the sufficiency score with a preset threshold, and make a judgment based on the current number of interaction iterations and the set maximum number of iterations. If the sufficiency score is greater than or equal to the set threshold, the currently retrieved context information is determined to be sufficient, and the process proceeds to step 11. If the sufficiency score is less than the set threshold, and the current number of interaction iterations is less than the set maximum number of iterations, the currently retrieved context information is determined to be insufficient, and the process proceeds to step 9. If the sufficiency score is less than the set threshold, and the current number of interaction iterations is greater than or equal to the set maximum number of iterations, the currently retrieved context set with the highest score is provided to the user, and the user is prompted that the information in the current knowledge base is incomplete or insufficient to form a completely definite suggestion, and the user is reminded to refer to the provided information with caution. Step 9: Use the large language model as a generative model and combine it with prompt words to generate targeted clarification questions, where the prompt words include the current query, the retrieved context set, and the calculated sufficiency score; Step 10: present the clarification question to the user, receive the user's answer, and then merge the answer information provided by the user with the current query to generate the next round of optimized query, and return to step 6; Step 11: Integrate the final confirmed query and the context set that is finally retrieved and confirmed to be used to form a complete generated prompt, and input the generated prompt into another large language model to generate the final answer to the user's rice pest and disease question.

2. The rice pest question-answering method based on circular retrieval according to claim 1, characterized in that: Step 1 specifically includes: Collect and organize unstructured rice pest and disease related text data from various channels, including publicly published scientific research papers on rice pathology, entomology, and pesticide science, technical reports and pest and disease forecasts issued by national or local plant protection stations, and prevention and control guidelines and variety resistance information issued by professional agricultural research institutions.

3. The rice pest question-answering method based on circular retrieval according to claim 2, characterized in that: Step 3 specifically includes: The text data of rice diseases and pests are segmented into blocks. For the chapters on a certain disease or pest in scientific research papers and technical reports, the titles are used as natural dividing points for segmentation. At the same time, a white list of professional terms in the field of rice diseases and pests is constructed. This white list is referred to during segmentation to ensure that text fragments containing these terms are not segmented incorrectly.

4. The rice pest question-answering method based on circular retrieval according to claim 1, characterized in that: The calculation methods for sufficiency score, keyword coverage, and content support are as follows: ； ； ； In the formula, Indicates keyword coverage. Indicates content support. represents the final context adequacy score, express and The length of the longest common subsequence, express Length, Represents the retrieved context, Indicates the optimized query keyword set. A set of keywords representing the retrieved context, and is the weight parameter.

5. The rice pest question-answering method based on circular retrieval according to claim 1, characterized in that: The optimized query is generated as follows: ,in represents the optimized query, Q represents the original question input by the user, and Trewrite represents the enhanced function.

Images