System, method, computer-readable storage medium, and computer program product for crop breeding question answering

Abstract

The present invention relates to a computer system using a computing model, and discloses a system, method, computer-readable storage medium, and computer program product for crop breeding question answering. A system for crop breeding question answering includes computing resources, a storage unit, and a large language model. The storage unit is configured to store a knowledge graph, which includes a crop breeding dataset. The large language model is configured to: receive input from a user; use predefined operators to search in the knowledge graph to determine the knowledge graph retrieval results; determine the credibility of the knowledge graph retrieval results; based on the credibility, selectively perform additional searches to determine the final retrieval results; and generate answers to crop breeding questions based on the crop breeding questions and the final retrieval results. The system according to the present invention overcomes the limitation of the large language model's lack of professional knowledge in the field of crop breeding, and improves the accuracy of the large language model in answering crop breeding questions.

Description

Technical Field

[0001] The present invention generally relates to computer systems utilizing computational models, and more particularly to systems, methods, computer-readable storage media, and computer program products for answering questions in crop breeding. Background Art

[0002] Plant biology research involves complex decision-making and requires integrating vast amounts of literature and multi-omics data. The rapid growth of biological data and publications poses a significant challenge to extracting knowledge and understanding from crop breeding. Large language models (LLMs) offer a promising solution to this challenge due to their exceptional performance in text generation.

[0003] However, while the general-purpose large language model excels in natural language processing tasks, it is a general-purpose AI product and lacks the ability to answer highly specialized crop breeding questions. For example, the general-purpose large language model's training corpus lacks specialized content from the crop breeding field, leading to misunderstandings of specialized terminology. For example, the general-purpose large language model's semantic parsing accuracy for rice-specific concepts (such as "leaf salt stress response during the tillering period") is less than 60%.

[0004] Furthermore, researchers are attempting to use knowledge graphs to answer specialized questions. While traditional knowledge graphs can represent gene function annotations, their reasoning capabilities are insufficient. Existing systems only support basic, rule-based queries and are unable to handle complex questions that require combining semantic and omics data (e.g., "Is a gene involved in drought stress response and has breeding potential?").

[0005] There is a need in the art for improved crop breeding question-answering technology on at least one of the aforementioned dimensions. Summary of the Invention

[0006] The present invention is provided to further improve crop breeding question-answering technology by utilizing large language models and knowledge graphs.

[0007] One aspect of the present invention provides a system for crop breeding question answering, comprising: computing resources; a storage unit, the storage unit being configured to store a knowledge graph, the knowledge graph comprising a crop breeding dataset; and a large language model, the large language model being configured to call the computing resources to: receive input from a user, the input comprising a crop breeding question; perform a search in the knowledge graph based on the crop breeding question using a predefined operator to determine a knowledge graph retrieval result, the predefined operator comprising an operation of formalized constraints required to retrieve the knowledge graph; determine the credibility of the knowledge graph retrieval result; based on the credibility, selectively perform an additional search to determine a final retrieval result; and generate an answer to the crop breeding question based on the crop breeding question and the final retrieval result.

[0008] In the system as described above, the large language model is configured to call the computing resources to: decompose the crop breeding problem into multiple sub-steps; perform the operation of at least one operator in each of the multiple sub-steps to obtain an intermediate retrieval result of each sub-step; and determine the knowledge graph retrieval result based on the intermediate retrieval result of the last sub-step in the multiple sub-steps.

[0009] In the system described above, the predefined operators are determined in the following manner: for each evaluation problem in the crop breeding test set, the large language model is used to decompose the evaluation problem into multiple sub-steps, each sub-step includes an operator; it is determined whether the operator list of the large language model covers the operators of the decomposed sub-steps of all evaluation problems, and the operator list includes the predefined operators; and the operator list is selectively updated according to the result of the determination; or for each evaluation problem in the crop breeding test set, the large language model is used to decompose the evaluation problem; for evaluation problems that fail to be decomposed, the expression of the operators in the operator list is selectively adjusted.

[0010] In the system as described above, the predefined operator includes at least one of knowledge graph query, arithmetic operation on data, logical operation on data, and result output.

[0011] In the system as described above, the large language model is configured to call the computing resources to: determine whether the credibility meets the requirements; in response to the credibility meeting the requirements, return the knowledge graph retrieval result as the final retrieval result; in response to the credibility not meeting the requirements, perform an additional search on the knowledge graph to determine the final retrieval result.

[0012] In the system as described above, in response to the credibility not meeting the requirement, the large language model is configured to call the computing resources to: decompose the crop breeding problem into a multi-level vocabulary, the multi-level vocabulary including a first-level vocabulary and a second-level vocabulary, the second-level vocabulary including the superordinate concepts of the first-level vocabulary; use the multi-level vocabulary to search in the knowledge graph to determine additional search results; and return the additional search results as the final search results.

[0013] In the system as described above, the large language model is configured to call the computing resource to: determine the complexity of the crop breeding problem; and in response to the complexity exceeding a complexity threshold, retrieve a predetermined number of relationships associated with the crop breeding problem in the knowledge graph, wherein the predetermined number is based on the length of the longest context of the large language model and the length of the text block mapped to the crop breeding dataset.

[0014] In the system as described above, the predetermined number of relationships includes the predetermined number of relationships with the highest priority, wherein the priority of the relationships in the knowledge graph is sorted based on the weight of the relationship determined when the knowledge graph is established, and the weight of the relationship reflects the closeness between the two entity nodes connected by the relationship in the knowledge graph.

[0015] In the system as described above, the input is a first input, and the large language model is configured to call the computing resource to: receive a second input from the user, the second input including a second question; determine the relevance of the second question to the field of crop breeding; and in response to the second question being irrelevant to the field of crop breeding, generate an answer to the second question based on the second question.

[0016] In the system as described above, the large language model is configured to call the computing resources to: divide the second question into semantic slots, and if the semantic slots are not related to the field of crop breeding, determine that the second question is not related to the field of crop breeding; or approximately match the dense features of the second question with the entity word list of the knowledge base of the large language model, and if the result of the approximate match is lower than a predetermined threshold, determine that the second question is not related to the field of crop breeding.

[0017] In the system as described above, the large language model is configured to: record historical messages of the user's conversation with the large language model, the historical messages including questions raised by the user; detect the relevance of the questions raised by the user to multiple aspects of the field of crop breeding; and in response to detecting that a threshold number of questions are related to a first aspect among the multiple aspects, increase the weight of the text block associated with the first aspect in the final retrieval result.

[0018] In the system as described above, the crop breeding dataset comprises multi-omics data.

[0019] In the system as described above, the crop breeding dataset is organized in the knowledge graph in the form of triples, wherein the triples include entities and relationships, and the entities of the triples of the multi-omics data include at least one of the following: genes, transcriptomes, proteins, events, gene annotations, traits, and environmental factors.

[0020] In the system as described above, the large language model is configured to: obtain literature related to crop breeding; extract triples of the crop breeding dataset from the literature; and construct the knowledge graph based on the triples of the crop breeding dataset, wherein the triples are mapped to text blocks in the literature.

[0021] In the system as described above, the triple includes a head entity and a tail entity, and the large language model is configured to: extract the gene name or variety name from the document as the head entity; when the head entity is the gene name, extract the breeding history, trait record or cultivation technology from the document as the tail entity; when the head entity is the variety name, extract the agronomic trait from the document as the tail entity.

[0022] In the system as described above, the knowledge graph is a first knowledge graph, and the system also includes a question-answering extraction model, which is configured to: extract a second knowledge graph from a crop breeding corpus, wherein the crop breeding corpus includes literature related to crop breeding; generate statements based on entities and relationships in the second knowledge graph; use the large language model to predict whether the statement is correct to generate a prediction result, wherein the prediction result indicates the confidence of the large language model in the statement; calculate the difference between the prediction result and the true value of the statement; in response to the difference being greater than an error threshold, determine the statement as a knowledge blind spot of the large language model; and extract a subgraph in the second knowledge graph based on the knowledge blind spot to generate a crop breeding question-answering sample training set for training the large language model.

[0023] Another aspect of the present invention provides a method for crop breeding question answering, comprising: calling computing resources by a large language model to perform the following steps: S101: receiving input from a user, the input comprising a crop breeding question; S102: using a predefined operator to search in a knowledge graph based on the crop breeding question to determine a knowledge graph retrieval result, the knowledge graph comprising a crop breeding dataset, the predefined operator comprising an operation of formal constraints required to retrieve the knowledge graph; S103: determining the credibility of the knowledge graph retrieval result; S104: based on the credibility, selectively performing an additional search to determine a final retrieval result; and S105: generating an answer to the crop breeding question based on the crop breeding question and the final retrieval result.

[0024] As described above, step S102 includes: S1021: decomposing the crop breeding problem into multiple sub-steps; S1022: performing at least one operator operation in each of the multiple sub-steps to obtain an intermediate retrieval result of each sub-step; and S1023: determining the knowledge graph retrieval result based on the intermediate retrieval result of the last sub-step in the multiple sub-steps.

[0025] In the method described above, the predefined operators include at least one of knowledge graph query, arithmetic operation on data, logical operation on data, and result output, and the predefined operators are determined in the following manner: for each evaluation problem in the crop breeding test set, the evaluation problem is decomposed into multiple sub-steps using the large language model, each sub-step includes an operator; judging whether the operator list of the large language model covers the operators of the decomposed sub-steps of all evaluation problems, the operator list including the predefined operators; and selectively updating the operator list based on the result of the judgment; or for each evaluation problem in the crop breeding test set, the evaluation problem is decomposed using the large language model; for evaluation problems that fail to be decomposed, the expression of the operators in the operator list is selectively adjusted.

[0026] As described above, step S102 includes: S1024: determining the complexity of the crop breeding problem; and S1025: in response to the complexity exceeding a complexity threshold, retrieving a predetermined number of relationships associated with the crop breeding problem in the knowledge graph, the predetermined number being based on the length of the longest context of the large language model and the length of the text block mapped to the crop breeding dataset, the predetermined number of relationships including the predetermined number of relationships with the highest priority, wherein the priority of the relationships in the knowledge graph is sorted based on the weight of the relationships determined when the knowledge graph is established, and the weight of the relationship reflects the closeness between two entity nodes connected by the relationship in the knowledge graph.

[0027] According to the method described above, the large language model calls computing resources to perform the following steps: S201: recording historical messages of the user's conversation with the large language model, the historical messages including questions raised by the user; S202: detecting the relevance of the questions raised by the user with multiple aspects of the field of crop breeding; and S203: in response to detecting that a threshold number of questions are related to a first aspect of the multiple aspects, increasing the weight of the text block associated with the first aspect in the final retrieval result.

[0028] According to the method described above, the crop breeding dataset includes multi-omics data, and the crop breeding dataset is organized in the knowledge graph in the form of triples, wherein the triples include entities and relationships, and the entities of the triples of the multi-omics data include at least one of the following: genes, transcriptomes, proteins, events, gene annotations, traits and environmental factors.

[0029] As described above, the large language model is configured to be used for: S401: obtaining literature related to crop breeding; S402: extracting triples of the crop breeding dataset from the literature, wherein the triples include a head entity and a tail entity, including: S4021: extracting a gene name or a variety name from the literature as a head entity; S4022: when the head entity is a gene name, extracting a breeding history, a trait record or a cultivation technique from the literature as a tail entity; S4023: when the head entity is a variety name, extracting agronomic traits from the literature as a tail entity; S403: constructing the knowledge graph based on the triples of the crop breeding dataset, wherein the triples are mapped to text blocks in the literature.

[0030] Another aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program implements the steps of any of the above methods when executed by a processor.

[0031] Another aspect of the present invention provides a computer program product, comprising a computer program, which implements the steps of any one of the above methods when executed by a processor.

[0032] The system and method according to the present invention overcome the limitation of the large language model's lack of professional knowledge in the field of crop breeding. By combining retrieval in the knowledge graph with the reasoning ability of the large language model, the accuracy of the large language model in answering crop breeding questions is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Various embodiments of the present invention are described with reference to the accompanying drawings.

[0034] Figure 1is a block diagram of a system for crop breeding question answering according to some embodiments of the present invention.

[0035] Figure 2 is a schematic diagram of a triple of a knowledge graph according to some embodiments of the present invention.

[0036] Figure 3 3 is a schematic diagram of a processing flow of a large language model generating an answer based on user input according to some embodiments of the present invention.

[0037] Figure 4 is a flow chart of a first method for crop breeding question answering according to some embodiments of the present invention.

[0038] Figure 5 is a flow chart of a first process associated with a first method for crop breeding question answering according to some embodiments of the present invention.

[0039] Figure 6 is a flow chart of a second process associated with the first method for crop breeding question answering according to some embodiments of the present invention.

[0040] Figure 7 is a flow chart of a third process associated with the first method for crop breeding question answering according to some embodiments of the present invention.

[0041] Figure 8 is a flow chart of a second method for crop breeding question answering according to some embodiments of the present invention.

[0042] Figure 9 is a flow chart of a third method for crop breeding question answering according to some embodiments of the present invention.

[0043] Figure 10 is a flowchart of a fourth method for crop breeding question answering according to some embodiments of the present invention.

[0044] Figure 11 is a flowchart of a fifth method for generating a training dataset according to some embodiments of the present invention.

[0045] Figure 12 is a block diagram of a computer-readable storage medium according to some embodiments of the present invention.

[0046] Figure 13 is a block diagram of a computer program product according to some embodiments of the present invention. DETAILED DESCRIPTION

[0047] In this application, ordinal numbers such as "first," "second," and "third" are used to distinguish different instances of the same object. The ordinal numbers such as "first," "second," and "third" do not indicate the relative order of the objects in time, space, ranking, or other aspects.

[0048] According to one aspect of the present invention, a system for question-answering in crop breeding is provided.

[0049] Figure 1 is a block diagram of a system 100 for crop breeding question answering according to some embodiments of the present invention.

[0050] System 100 may be a local or remote computer, server, etc. System 100 may include computing resources 110, storage unit 120, and a large language model 130. Storage unit 120 may be configured to store a knowledge graph, which may include a crop breeding dataset.

[0051] In some embodiments, the computing resources 110 may include a central processing unit (CPU), a graphics processing unit (GPU), and various other processing units or cores (e.g., arithmetic logic unit, integer unit, floating point unit, tensor unit, ray tracing core, etc.).

[0052] The large language model 130 can be configured to call the computing resources 110 to perform corresponding operations. The large language model 130 can be configured to receive input from a user, the input including a crop breeding question. The large language model 130 can be configured to use predefined operators to search in the knowledge graph based on the crop breeding question to determine the knowledge graph retrieval results. The predefined operators include operations on formal constraints required to retrieve the knowledge graph. The large language model 130 can be configured to determine the credibility of the knowledge graph retrieval results. The large language model 130 can be configured to selectively perform additional searches based on the credibility to determine the final retrieval results. The large language model 130 can be configured to generate answers to the crop breeding questions based on the crop breeding questions and the final retrieval results.

[0053] In some embodiments, the credibility of knowledge graph search results can be determined based on the results themselves. For example, the large language model 130 can compare the results with data from an open-source seed industry evaluation set to determine the credibility of the results. For example, the open-source seed industry evaluation set can include at least one of multiple-choice questions, fill-in-the-blank questions, and essay questions, each of which includes an objective answer. The large language model 130 can compare the knowledge graph search results with the objective answers and determine the credibility of the knowledge graph search results based on the comparison results (e.g., the similarity between the results and the objective answers).

[0054] In some embodiments, the credibility of a knowledge graph search result can be determined based on the process that generated the result. For example, the large language model 130 can determine whether the credibility of a knowledge graph search result meets requirements based on the success rate of decomposing a crop breeding problem into predefined operators and executing the operators.

[0055] This paper proposes a large language model for the field of crop breeding. By combining knowledge graph retrieval with the reasoning ability of the large language model, it realizes the dynamic integration of the text data of the large language model itself and the data in the knowledge graph. It enables the system to dynamically fuse multi-source data (such as the text data of the large language model and the crop breeding data in the knowledge graph) according to user queries, and generate comprehensive and accurate responses, thus breaking through the limitations of existing general large language models in dealing with professional problems in crop breeding.

[0056] In some embodiments, crop breeding datasets may include multi-omics data. Multi-omics data refers to a collection of biological data obtained through various omics technologies (such as genomics, transcriptomics, proteomics, and metabolomics). Integrating crop breeding datasets containing multi-omics data (such as genomics, transcriptomics, and proteomics) through knowledge graphs provides richer biological data for large language models to answer crop breeding-related questions.

[0057] In some embodiments, crop breeding datasets can be organized in a knowledge graph as triples. Triples consist of entities and relationships, where entities can also be referred to as nodes and relationships can also be referred to as edges. Entities in a triple of multi-omics data can include at least one of the following: genes, transcriptomes, proteins, events, gene annotations, traits, and environmental factors.

[0058] Figure 2 Figure 2 is a schematic diagram of a knowledge graph triple according to some embodiments of the present invention. Circles indicate entities in the triple. In some embodiments, the entities in the triple may include gene, transcript, protein, event, gene annotation (kegg, go), trait (pheno), and / or environmental factor (env_factor). The number next to the circle indicates the number of instances of the corresponding entity. The lines connecting the circles indicate the relationship between the entities.

[0059] It should be understood that Figure 2 It is a statistical diagram of triples of the knowledge graph constructed according to some embodiments of the present invention. For a single triple, the triple itself exists in the form of one entity connected to another entity.

[0060] A triple can be represented in the form of (head entity, relation, tail entity). The following is an example of a triple in a knowledge graph:

[0061] (1) Gene nodes, each piece of information contains the unique identifier of the gene in the map. Gene nodes can be related to GO annotations, Kegg annotations, genes, traits, and transcriptomes:

[0062] Relationship with GO: (AGIS_Os06g001280, in go, GO:0003674);

[0063] Relationship with Kegg: (AGIS_Os07g036840, in kegg, K01126);

[0064] Whether the two genes form an upstream and downstream relationship: (AGIS_Os09g044250, near, AGIS_Os09g044260);

[0065] Whether two genes are co-expressed: (AGIS_Os11g026770, coexpression, AGIS_Os03g050730);

[0066] Relationship between genes and environmental factors: (AGIS_Os05g024200, env_factor_phenotype, Water-related factors);

[0067] Relationship between gene and trait: (AGIS_Os06g001280, Phenotype, interact with chloroplast);

[0068] Relationship between gene and transcriptome: (AGIS_Os02g050860, transcript, transcript of AGIS_Os02g050860);

[0069] (2) Transcriptome node: Each gene contains a unique transcriptome, which can be related to genes, proteins, and transcriptome events:

[0070] Relationship between gene and transcriptome: (AGIS_Os02g050860, transcript, transcript of AGIS_Os02g050860);

[0071] Relationship between transcriptome and protein: (transcript of AGIS_Os05g033520, translate, protein of AGIS_Os05g033520);

[0072] Relationship between transcriptome events and transcriptome: (transcriptome_event_0, involves_transcript, transcript of AGIS_Os01g002320);

[0073] (3) Protein nodes can be related to transcriptomes, proteins, and protein events:

[0074] Relationship between transcriptome and protein: (transcript of AGIS_Os05g033520, translate, protein of AGIS_Os05g033520);

[0075] Protein-protein interaction (PPI): (protein of AGIS_Os01g055060, Protein_protein_interaction, protein of AGIS_Os02g053780);

[0076] Relationship between protein and protein event: (proteome_event_0, involves_protein, proteinof AGIS_Os05g031310);

[0077] (4) Event nodes, including transcriptome events and protein events, can be related to transcriptomes and proteins. In addition, tissue type and growth stage can be regarded as attributes of the node:

[0078] Relationship between transcriptome events and transcriptome: (transcriptome_event_0, involves_transcript, transcript of AGIS_Os01g002320);

[0079] Protein-protein event relationship: (proteome_event_0, involves_protein, proteinof AGIS_Os05g031310);

[0080] Growth stage of protein event: (proteome_event_0, observed_in_stage, SeedlingStage);

[0081] Tissue type of protein event: (proteome_event_2, observed_in_TissuesOrgans,Root);

[0082] Tissue type of transcriptome event: (transcriptome_event_200919, observed_in_TissuesOrgans, Leaf);

[0083] (5) Kegg annotation nodes can be related to genes:

[0084] Relationship between gene and Kegg: (AGIS_Os07g036840, in kegg, K01126);

[0085] (6) GO annotation nodes can be related to genes:

[0086] Relationship between gene and GO: (AGIS_Os06g001280, in go, GO:0003674);

[0087] (7) Trait nodes can be related to genes and environmental factors:

[0088] Relationship between gene and trait: (AGIS_Os06g001380, Phenotype, affect ethylenehomeostasis);

[0089] Relationship between genes and environmental factors: (AGIS_Os05g024200, env_factor_phenotype, Water-related factors);

[0090] (8) Environmental factors:

[0091] Relationship between genes and environmental factors: (AGIS_Os05g024200, env_factor_phenotype, Water-related factors).

[0092] In some embodiments, the entities of the triplet may include gene name, variety name, breeding history, trait record, cultivation technology, agronomic traits, etc.

[0093] It should be understood that the above are merely examples of triples in the knowledge graph and are not exhaustive, and the scope of the present invention is not limited thereto.

[0094] Some embodiments of the present invention use triple structures to represent complex associations between multi-omics data, providing a new paradigm for the structured representation of crop breeding data, making it easier for large language models to retrieve and generate information related to crop breeding problems from massive data, and improving data processing efficiency.

[0095] The knowledge graph can also be mapped to text chunks. Text chunks are stored in a corpus generated based on content such as documents and images. Each text chunk has an approximate length, referred to as the chunk length. Each entity and relationship in a triple in the knowledge graph is mapped to a text chunk in the corpus. The chunk identifier corresponding to the text chunk can be recorded in each entity and relationship, allowing indexing to the text chunk in the corpus using the chunk identifier recorded in the entity and relationship.

[0096] In some embodiments, the predefined operators may include at least one of knowledge graph query, arithmetic operation on data, logical operation on data, and result output.

[0097] Table 1 below is an example table of operators. Arithmetic operations on data can include accumulation, etc., and logical operations on data can include counting, comparison, etc. Operators are used to constrain the process of retrieving knowledge graphs based on formal methods. This means that the operations involved in retrieving knowledge graphs are divided into a limited number of specific sub-steps based on the nature of the operations. Each type of sub-step can be represented and executed through "formal expression" (i.e., an abstract method that uses symbols, rules, and mathematical structures to unambiguously and precisely describe complex systems or concepts. Its core is to replace the ambiguity of natural language with strictly defined syntax and semantics). This constraint can make the retrieval process highly interpretable. The clear parameter definition of each operator can reduce the context length requirements of large language models, thereby reducing the demand for video memory.

[0098]

[0099] In some embodiments, the large language model 130 may be configured to decompose the crop breeding problem into multiple sub-steps. The large language model 130 may be configured to perform at least one operator in each of the multiple sub-steps to obtain an intermediate search result for each sub-step. The large language model 130 may be configured to determine a knowledge graph search result based on the intermediate search result of the last sub-step in the multiple sub-steps.

[0100] The following is an example of searching in a knowledge graph based on a crop breeding problem using predefined operators.

[0101] ===========;

[0102] json "query": "What is the parental relationship between Huang Fengzhan and Feng Xiuzhan?",

[0103] "answer": [

[0104] {"step":"Query the parent relationship between Huang Fengzhan and Feng Xiuzhan","action":"get_spo(s=s1:Rice[Huang Fengzhan], p=p1:Parent relationship, o=o1:Rice[Feng Xiuzhan])"},

[0105] {"step":"Query parent relationship","action":"get(p1)"} ]

[0107] ===========;

[0108] First, the large language model 130 uses predefined operators to decompose the crop breeding question, "What is the parental relationship between Huang Fengzhan and Feng Xiuzhan?", into two steps. It then executes the operators in each of these two steps, using variables (e.g., s1, p1, o1) to record the intermediate results of the first step. Finally, in the second (and final) step, it obtains the knowledge graph search results for the crop breeding question.

[0109] The following are other examples of searching in the knowledge graph based on crop breeding questions using predefined operators.

[0110] 1) Crop breeding problems involving simple statistics:

[0111] ===========;

[0112] Plain Text

[0113] query: How many AGIS genes are involved in drought stress?

[0114] reasoning: [

[0116] {

[0117] "step": "Query AGIS genes involved in drought stress",

[0118] "action": "get_spo(s=s1:gene[AGIS], p=p1:involved stress type, o=o1:stress type[drought])"

[0119] },

[0120] {

[0121] "step": "Count the number of genes that meet the conditions",

[0122] "action": "count(s1)"

[0123] } ]

[0125] response:

[0126] According to the reference documents provided, the AGIS genes involved in drought stress are as follows:

[0127] 1. AGIS_Os05g040410: Differentially expressed in different species and tissues and organs:

[0128] - In the leaves of variety Nagina-22, it showed an upregulation (Up).

[0129] - In the grain embryos of variety Zhonghua11, it showed upregulation (Up).

[0130] - In the glumes of variety YIL100, it showed an up-regulation (Up).

[0131] - In the roots of Rice 9311, it showed an upregulation (Up).

[0132] 2. AGIS_Os10g034460: In the glumes of variety YIL100, it showed upregulation (Up).

[0133] - In the roots of Rice 9311, it showed an upregulation (Up).

[0134] In summary, there are two AGIS genes involved in drought stress.

[0135] ===========;

[0136] 2) Crop breeding issues involving comparison of different genes:

[0137] ===========;

[0138] Plain Text

[0139] query: How many types of similar environmental responsiveness are there between AGIS_Os05g040410 and AGIS_Os06g035130?

[0140] logic form: [

[0142] {"step":"Query the environmental responsiveness of AGIS_Os05g040410","action":"get_spo(s=s1:gene[AGIS_Os05g040410], p=p1:environmental responsiveness, o=o1)"},

[0143] {"step":"Query the environmental responsiveness of AGIS_Os06g035130","action":"get_spo(s=s2:gene[AGIS_Os06g035130], p=p2:environmental responsiveness, o=o2)"},

[0144] {"step":"Calculate the number of similar environmental responsiveness of two genes","action":"compare(set=[o1, o2], op=equal)->equal_count"},

[0145] {"step":"Count the number of similar environmental responsiveness types","action":"count(equal_count)->count_result"} ]

[0147] Intermediate results:

[0148] step1

[0149] According to the reference document, the environmental responsiveness types of AGIS_Os05g040410 include the following:

[0150] 1. Cold Stress Response

[0151] 2. Cadmium Stress Responses and Tolerance in Rice

[0152] 3. Early Chilling Stress Response in Rice

[0153] 4. Rice Drought Stress Response

[0154] 5. Rice Heat Stress Response

[0155] 6. Low-Level Gamma Radiation

[0156] A total of 6 different environmental responsiveness.

[0157] step2

[0158] According to the reference document, the environmental responsiveness types of AGIS_Os06g035130 include the following:

[0159] 1. Cadmium Stress Responses and Tolerance in Rice

[0160] 2. Salt Tolerance in Rice

[0161] 3. Rice Drought Stress Response

[0162] 4. Oxidative Stress Response in Rice

[0163] 5. Strigolactone Stress in Rice

[0164] 6. Aerobic Adaptation in Rice

[0165] 7. ABA Signaling and Biosynthesis

[0166] A total of 7 different environmental responsiveness.

[0167] step3

[0168] According to the information provided, AGIS_Os05g040410 has 6 types of environmental responsiveness, while AGIS_Os06g035130 has 7 types of environmental responsiveness. We need to find out how many types of environmental responsiveness they have in common.

[0169] First, list all environmental responsiveness categories of AGIS_Os05g040410:

[0170] 1. Cold stress response

[0171] 2. Rice cadmium stress response

[0172] 3. Early cold stress response in rice

[0173] 4. Rice drought stress response

[0174] 5. Rice heat stress response

[0175] 6. Low-level gamma radiation response

[0176] Then, list all environmental responsiveness categories of AGIS_Os06g035130:

[0177] 1. Rice cadmium stress response

[0178] 2. Rice salt stress response

[0179] 3. Rice drought stress response

[0180] 4. Rice Oxidative Stress Response

[0181] 5. Response of rice to gibberellin stress

[0182] 6. Adaptive response of rice shoot roots

[0183] 7. Rice and Arabidopsis receptor-like kinase gene response

[0184] Comparing these two lists, we can see that they have two common types of environmental responsiveness:

[0185] 1. Rice cadmium stress response

[0186] 2. Rice drought stress response

[0187] Therefore, the number of similar environmental responsiveness types between AGIS_Os05g040410 and AGIS_Os06g035130 is 2.

[0188] Final reply:

[0189] How many types of environmental responsiveness are similar between AGIS_Os05g040410 and AGIS_Os06g035130?

[0190] According to the reference information, the number of similar environmental responsiveness types between AGIS_Os05g040410 and AGIS_Os06g035130 is 2. These common environmental responsiveness types include:

[0191] 1. Rice cadmium stress response

[0192] 2. Rice drought stress response

[0193] This conclusion was reached by comparing the environmental responsiveness lists of the two genes.

[0194] ===========;

[0195] Traditional large language models are unfamiliar with professional terms (such as rice varieties) and concepts in the field of crop breeding. When encountering queries containing unfamiliar professional terms and concepts, the retrieval of knowledge graphs degenerates into fuzzy matching, which greatly reduces the credibility of the query results.

[0196] As the above examples demonstrate, by using predefined operators to search the knowledge graph for crop breeding problems, and executing searches at a finer granularity (each sub-step), the entire search process becomes more interpretable, increasing the probability of matching specialized terms and concepts during the query, thereby enhancing the credibility of the search results. Furthermore, each operator has clear input parameters, which reduces the maximum context length requirement for large language models, thereby reducing the memory requirements for hardware execution.

[0197] During the search process, whether the operator list, including predefined operators, includes all the operators required to answer the question is crucial for the success of the search. To improve operator support for crop breeding questions, the questions in the crop breeding test set can be broken down to improve the coverage of the operator list. The crop breeding test set can be a manually generated question-answer set used to evaluate the performance of large language models in answering questions in the crop breeding field. The specific content of the crop breeding test set is detailed below.

[0198] In some embodiments, predefined operators can be determined in the following manner. First, for each evaluation problem in the crop breeding test set, the large language model 130 is used to decompose the evaluation problem into multiple sub-steps, each of which includes an operator. Then, a determination is made as to whether the operator list of the large language model 130 covers all operators of the decomposed sub-steps of the evaluation problem, and whether the operator list includes predefined operators. Finally, the operator list is selectively updated based on the determination result.

[0199] In some embodiments, the predefined operators may be determined in the following manner: For each evaluation question in the crop breeding test set, the evaluation question is decomposed using the large language model 130. For evaluation questions that fail to be decomposed, the expression of the operators in the operator list is selectively adjusted.

[0200] By increasing the coverage of crop breeding questions in the operator list, it helps to improve the accuracy and comprehensiveness of retrieval results and provides a basis for large language models to generate better answers.

[0201] Additionally or alternatively, the prompts of the large language model may be optimized to improve the success rate of decomposing the large language model into sub-steps.

[0202] In some embodiments, successful decomposition examples for crop breeding problems may be added to the prompts of the large language model to guide the large language model to perform decomposition in a targeted manner.

[0203] Here is an example of a successful decomposition for a crop breeding problem:

[0204] ===========;

[0205] "query": "What is the parental relationship between Huang Fengzhan and Feng Xiuzhan?",

[0206] "answer": [{"step":"Query the parent relationship between Huang Fengzhan and Feng Xiuzhan","action":"get_spo(s=s1:Rice[Huang Fengzhan], p=p1:Parent relationship, o=o1:Rice[Feng Xiuzhan])"},{"step":"Query the parent relationship","action":"get(p1)"}]

[0207] ===========;

[0208] In some embodiments, a requirement for question decomposition may be added to the prompts of the large language model.

[0209] The following is an example of the requirements for problem decomposition:

[0210] ===========;

[0211] ## Output requirements

[0212] - The output is a list, each element of the list contains a step and an action. The step is the description of the sub-problem, and the action is the function to be executed.

[0213] - Each sub-problem can only execute 1 function, not multiple

[0214] - You don't nest a function into another function's parameters

[0215] - You will not only output natural language, if there is no suitable function to solve the problem, directly reply "no suitable function"

[0216] ## Notes

[0217] - You only need to break down the problem into sub-steps from an NLP perspective, **no need to answer the question itself**

[0218] ===========;

[0219] In some cases, when searching a knowledge graph using predefined operators, the resulting search results may not meet the requirements. For example, for rare crop breeding problems, there may not be a corresponding predefined operator, resulting in search results that do not meet the requirements. Therefore, to improve the accuracy of search results, it is necessary to determine whether the search results meet the requirements.

[0220] In some embodiments, the large language model 130 may be configured to determine whether the credibility of the knowledge graph retrieval results meets the requirements.

[0221] In some embodiments, in response to the credibility meeting the requirement, the large language model 130 can be configured to return the knowledge graph search result as the final search result. In this case, the search result obtained using the predefined operator is the final search result.

[0222] In some embodiments, in response to the confidence level not meeting the requirement, the large language model 130 may be configured to perform an additional search on the knowledge graph to determine a final search result.

[0223] For example, the large language model 130 may determine whether the credibility of the search results meets the requirements based on the following prompts.

[0224] ===========;

[0225] yaml

[0226] You are a Chinese and English examiner, good at analyzing whether students can get the answers to the questions based on their knowledge, and giving YES or NO accordingly.

[0227] ## Task

[0228] Please read the test questions and student evidence carefully to determine whether the student evidence can answer the test questions.

[0229] ## Output format requirements

[0230] - If the student's evidence is insufficient, output NO

[0231] - If the student's basis contains or can deduce the answer to the test question, output YES

[0232] - Before giving the final YES / NO, you will explain why you made this judgment

[0233] ## Notes

[0234] - When checking students' answers, you will notice the difference in units of expression, for example, 1 kg is 2 jin

[0235] ## Student Answer Example

[0236] - Output NO after interpretation: "Cannot be determined. The information in options A, B, C, and D does not match the information provided in the literature."

[0237] - Output NO after explanation: "Based on the available information, we cannot determine the female and male parents of Koshihikari. However, the literature mentions that Koshihikari was used as a parent in hybridization with other conventional varieties (lines)."

[0238] - Output YES after explanation: "According to the provided information, Koshihikari's parents are Kinki 34 (♀) and Hokuriku 4 (♂)"

[0239] ## Exam Questions

[0240] {input_query}

[0241] ## Student basis

[0242] ```txt

[0243] {input_evidence}

[0244] ```

[0245] ===========;

[0246] In some embodiments, the credibility of the knowledge graph search result can be determined based on the result itself. The manner of determining the credibility of the knowledge graph search result based on the result itself is as described above for the large language model 130.

[0247] In some embodiments, the credibility of knowledge graph retrieval results can be determined based on the process that generated the results. For example, the large language model 130 can determine whether the credibility of the retrieval results meets requirements based on the success rate of decomposing a crop breeding problem into predefined operators and executing these operators. For example, the pass@k metric can be used to evaluate the success rate of the large language model 130 in decomposing and executing the crop breeding problem. For example, pass@k = number of sub-steps passed / total number of sub-steps. If this metric is greater than a predetermined threshold, the credibility is met. Otherwise, the credibility is not met.

[0248] By determining whether the credibility meets the requirements and responding differently according to different situations, the large language model can obtain more accurate retrieval results as the basis for generating the final answer.

[0249] When the confidence level does not meet the requirements, additional searches such as Graph-based Retrieval-Augmented Generation (GraphRAG) can be performed as a fallback strategy to provide basic answers to crop breeding questions.

[0250] In some embodiments, in response to the confidence level not meeting the requirements, the large language model 130 may be configured to decompose the crop breeding problem into a multi-level vocabulary. The multi-level vocabulary may include first-level vocabulary and second-level vocabulary, where the second-level vocabulary may include hypernyms of the first-level vocabulary. The large language model 130 may be configured to search the knowledge graph using the multi-level vocabulary to determine additional search results. The large language model 130 may be configured to return the additional search results as the final search results.

[0251] As an example, first, the crop breeding problem input by the user is decomposed into multi-level vocabulary. The following is an example of the decomposition process:

[0252] ===========;

[0253] Plain Text

[0254] query: "How many years has Zhejiang Radiation 802 been planted?"

[0255] low_level: [Zhejiang Radiation 802, year]

[0256] high_level: [rice variety, planting time]

[0257] ===========;

[0258] Among them, rice variety is the superordinate concept of Zhejiang Fu 802, and planting time is the superordinate concept of year.

[0259] Then, the first-level and second-level vocabulary are used to retrieve the associated entities and relations in the knowledge graph and mapped to the text blocks corresponding to the retrieved entities and relations. Therefore, the retrieved entities, relations, and text blocks can be used as the final retrieval results.

[0260] Therefore, when the credibility of the initial search results does not meet the requirements, by performing additional searches, it can be ensured that the search results meet the basic requirements.

[0261] When the crop breeding questions input by the user are more complex, the large language model may retrieve several entities, relationships and corresponding text blocks in the knowledge graph, but there is an upper limit to the context length that the large language model can handle. Therefore, the retrieval process needs to be adjusted according to the complexity of the question.

[0262] In some embodiments, large language model 130 may be configured to determine the complexity of a crop breeding question. In response to the complexity exceeding a complexity threshold, large language model 130 is configured to retrieve a predetermined number of relationships associated with the crop breeding question from the knowledge graph. The predetermined number is based on the length of the longest context of large language model 130 and the length of the text block mapped to the crop breeding dataset.

[0263] As an example, the Measure of Textual Lexical Diversity (MTLD) metric can be used to measure the complexity of a problem. It evaluates the lexical richness of a text by calculating the frequency and distribution of different words in the text. The calculation formula is:

[0264] (Formula 1)

[0265] in, is the total number of words in the text, is the number of distinct words in the text, It is The frequency of different words in the text.

[0266] The predetermined number can be calculated as follows. For example, if the length of the largest language model's longest context is llm_max_length, and the length of the longest single text block mapped to the knowledge graph is chunk_length, then the predetermined number top_k = llm_max_length / chunk_length, corresponding to the maximum number of text blocks that the large language model's context can accommodate. When there are multiple candidate relationships for a crop breeding problem, limiting the number of retrievals can reduce the amount of data processed by the large language model and improve processing efficiency.

[0267] In some embodiments, the predetermined number of relationships may include a predetermined number of relationships with the highest priority. The priorities of the relationships in the knowledge graph may be sorted based on the weights of the relationships determined when the knowledge graph is created. The weight of a relationship reflects the closeness between the two entity nodes connected by the relationship in the knowledge graph. For example, when creating the knowledge graph, the closeness between entities may be scored by a large language model to determine the weight of the corresponding relationship.

[0268] The process of performing corresponding searches based on credibility will be discussed below. Figure 3 Further description.

[0269] During a conversation between a user and a large language model, their questions may be broad or focused on a specific field, or even a specific aspect within a field. When a user's questions repeatedly touch on a specific aspect within a field, they likely possess expertise in that area. If the large language model provides answers that include more relevant content, this helps improve user satisfaction with the answers.

[0270] In some embodiments, large language model 130 may be configured to record historical messages of a user's conversation with large language model 130, including questions posed by the user. Large language model 130 may be configured to detect the relevance of questions posed by the user to multiple aspects of the field of crop breeding. In response to detecting that a threshold number of questions are related to a first aspect among the multiple aspects, large language model 130 may be configured to increase the weight of a text block associated with the first aspect in the final search results.

[0271] As an example, during the interaction between the user and the large language model 130, historical messages of the conversation may be recorded. The specific expression may be in ChatML format, for example:

[0272] ===========;

[0273] Plain Text [

[0275] {"role":"user","content":"Hello"},

[0276] {"role":"assistant","content":"Hello! I'm Fengdeng. I can provide you with information and services related to seed industry, agricultural technology, crop management, etc. How can I help you?"},

[0277] {"role":"user","content":"What is Zhejiang Radiation 802?"},

[0278] {"role":"assistant","content":"Zhefu 802 is a rice mutant variety developed through collaborative breeding between the Institute of Nuclear Agronomy of Zhejiang Agricultural University and the Yuhang County Agricultural Science Research Institute."},

[0279] {"role":"user","content":"What is Yebai?"} ]

[0281] ===========;

[0282] From the above, we can see that the current question "What is Yebai?" and the previous question "What is Zhefu 802?" are both related to rice varieties, and it can be assumed that the user has knowledge of rice breeding. Therefore, in the subsequent re-ranking calculation of the text blocks in the search results provided as context to the large language model 130, the text blocks related to rice can be weighted to reduce the number of responses unrelated to rice breeding. For example, a specific implementation is as follows:

[0283] ===========;

[0284] Python

[0285] from typing import List

[0286] class Chunk:

[0287] def __init__(self, label: str, score: float):

[0288] self.label = label

[0289] self.score = score

[0290] class Query:

[0291] def __init__(self, label: str, default_bias: float = 1.5):

[0292] self.label = label

[0293] self.default_bias = default_bias

[0294] def rerank(chunks: List[Chunk], query: Query) -> List[Chunk]:

[0295] # Traverse each block and weight the blocks that match the user's domain knowledge

[0296] for chunk in chunks:

[0297] if chunk.label == query.label:

[0298] chunk.score *= query.default_bias # weighted processing

[0299] # Reorder blocks by their scores (higher scores first)

[0300] chunks.sort(key=lambda x: x.score, reverse=True)

[0301] return chunks

[0302] ===========;

[0303] When it is identified that the user has professional knowledge in a related field, by focusing the answer on that field, the closeness between the answer and the question can be improved, thereby enhancing the user experience.

[0304] Therefore, in the process of interacting with users, for crop breeding issues, the large language model of some embodiments of the present invention utilizes a knowledge graph containing professional information combined with its reasoning ability to provide more accurate answers.

[0305] In order to construct a knowledge graph related to crop breeding, it is necessary to extract crop breeding related data from crop breeding related literature and extract triples from the crop breeding related data. However, crop breeding related data has the following characteristics:

[0306] (1) Multimodal heterogeneity: data containing multiple modalities such as gene sequencing data, phenotypic information, environmental data, and literature text data;

[0307] (2) Dynamic temporal nature: multi-stage experimental data covering the entire crop growth cycle (germination to maturity);

[0308] (3) Complex association: There are nonlinear interactions between genotype, phenotype and environment (G×E×P), involving more than 200 agronomic trait indicators;

[0309] (4) Domain specificity: including crop-specific gene naming systems (e.g., Os01g0705600).

[0310] Due to the multidimensional characteristics of crop breeding-related data, existing regular / fuzzy search methods and natural language processing (NLP) methods cannot generate complete results when extracting triples.

[0311] As an example, the triple extraction task is as follows:

[0312] ===========;

[0313] Bash

[0314] # Review report text

[0315] Yangdao No. 6 is a conventional indica rice variety. It has a short, sturdy seedling stage, good growth, and medium tillering. It has a plant height of 110 cm, a 1,000-grain weight of 30 g, and a full growth period of approximately 156 days.

[0316] # Document description text

[0317] "OsPPKL1 inhibits the expression of OsCKX2 in panicles, leading to an increase in the number of grains per panicle (p<0.01)"

[0318] ==========;

[0319] For the regular / fuzzy search method, the results are as follows:

[0320] ==========;

[0321] Bash

[0322] # Regular results of the approval report text

[0323] [('Yangdao 6', 'Plant height', '110 cm'),

[0324] ('Yangdao 6', '1000-grain weight', '30 g'},

[0325] ('Yangdao 6', 'Full growth period', '156 days')]

[0326] # Regular results of the literature description text

[0327] [('OsPPKL1', 'Inhibits', 'OsCKX2')]

[0328] ==========;

[0329] For the regular / fuzzy search method, various situations need to be considered during the search process, such as whether the variety name should include the character "number" (Yangdao 6 / Yangdao 6), and whether the units of various agronomic traits are abbreviated in letters (centimeter / cm, gram / g, day / d). In addition, as can be seen from the results, the regular / fuzzy search method cannot capture unstructured descriptions such as "medium tillering ability", and completely loses information such as "increase in the number of grains per panicle" and significance (p-value). Furthermore, the regular / fuzzy search method cannot automatically associate gene IDs and gene symbols in the text. For example, the gene OsCKX2 cannot be corresponding to Os01g0197700 and Os01g0197700, so only searching for gene IDs will lead to missed detections.

[0330] For the natural language processing method, the results are as follows:

[0331] ==========;

[0332] python

[0333] # NLP results of the approval report text

[0334] [('Yangdao 6', 'Plant height', '110 cm'),

[0335] ('Yangdao 6', '1000-grain weight', '30 g'},

[0336] ('Yangdao No. 6', 'Full Growth Period', '156 days'),

[0337] ("Yangdao No. 6", "Type", "Indica conventional rice"),

[0338] ("Yangdao No. 6", "has characteristics", "short and strong in seedling stage"),

[0339] ("Yangdao No. 6", "has characteristics", "medium tillering")]

[0340] # Document description text nlp results

[0341] [("OsPPKL1", "Negative regulation", "OsCKX2"), ("OsCKX2", "Effect", "Number of grains per spike"), ("The above relationship", "Significance", "p<0.01")]

[0342] ===========;

[0343] Natural language processing methods rely on syntactic analysis and rule templates and can capture some unstructured relationships. However, such methods usually rely on predefined sets of entities and relationships, have strong prior constraints, require labeled samples for agronomic traits, and rely on manual rule verification for specifications.

[0344] Therefore, some embodiments of the present invention propose to use a large language model to construct a knowledge graph.

[0345] In some embodiments, the large language model 130 may be configured to obtain literature related to crop breeding. The large language model 130 is configured to extract triples of a crop breeding dataset from the literature. The large language model 130 is configured to construct a knowledge graph based on the triples of the crop breeding dataset, where the triples are mapped to text blocks in the literature.

[0346] For example, a search can be performed in a database in Chinese or English based on crop breeding related terms to obtain literature. The large language model 130 can then extract triples based on the prompts to construct a knowledge graph. The following is an example of a prompt:

[0347] ===========;

[0348] Plain Text

[0349] ## Character Settings:

[0350] You are an agricultural informatics expert responsible for extracting structured knowledge from multi-source breeding data. You need to combine knowledge of molecular biology, agronomic trait standards, and breeding practices.

[0351] Mission Objectives:

[0352] Build a high-quality knowledge graph of crop breeding data based on the provided text

[0353] ## Restrictions

[0354] 1. Ambiguous terms such as "a gene" are prohibited. Unknown genes should be marked as "UNK-GENE-[serial number]"

[0355] 2. Yield-related traits must indicate the measurement standards

[0356] 3. When the text mentions certain unquantified conditions, it will be associated with the evaluation level. For example, when the text mentions "drought conditions", it will automatically be associated with the drought level. 4. ......

[0358] ## Thinking process:

[0359] 1. Understand the text content;

[0360] 2. Locate and build each set of associations in turn;

[0361] 3. When constructing each set of associations, the knowledge information implicit in the associations

[0362] 4. Will return standard json structure data

[0363] ## Text content:

[0364] {}

[0365] ===========;

[0366] For the large language model, the results of extracting triples are as follows:

[0367] ===========;

[0368] Python

[0369] # Review report text LLM results [

[0371] ("Yangdao No. 6", "Classification", "Indica conventional rice"),

[0372] ("Yangdao No. 6", "plant height", "110 cm"),

[0373] ("Yangdao No. 6", "1000-grain weight", "30g"),

[0374] ("Yangdao No. 6", "Growing period", "156 days"),

[0375] ("Yangdao No. 6", "Performance traits", "Stunted and strong at the seedling stage"),

[0376] ("Yangdao No. 6", "Tillering Characteristics", "Medium") ]

[0378] # Document description text LLM results

[0379] {"core_relation":

[0380] {"subject": "OsPPKL1",

[0381] "action": "Suppress",

[0382] "object": "OsCKX2",

[0383] "effect":

[0384] {"trait": "Number of grains per ear",

[0385] "direction": "Increase",

[0386] "tissue": "spike"},

[0387] "evidence":

[0388] {"p_value": 0.01,

[0389] "experiment_type": "Gene Expression Analysis"}

[0390] }

[0391] }

[0392] ===========;

[0393] By using large language models to extract triplets and construct knowledge graphs, implicit knowledge can be automatically associated, the biological significance of regulatory relationships can be understood (for example, inhibition corresponds to negative regulation), and regulatory chains can be linked (for example, from OsPPKL1 to OsCKX2 and then to phenotypes). Therefore, by leveraging the information extraction capabilities of large language models, open-source mining of crop breeding knowledge can be achieved, breaking through the pre-set limitations of traditional methods on "entity-relationship" categories and improving the flexibility and coverage of knowledge acquisition.

[0394] The large language model 130 has the following capabilities when building a knowledge graph:

[0395] (1) Complex relationship modeling capability: capturing complex relationship descriptions in the literature;

[0396] (2) Cross-literature knowledge fusion: automatically associate equivalent descriptions in different documents (e.g., "thousand-grain weight" ≈ "TKW" ≈ "1000-grain weight", OsGSL1 corresponds to LOC_Os01g12345);

[0397] (3) Self-check of agronomic trait specifications: whether the traits mapped by genes meet the agronomic trait specifications;

[0398] (4) Identification of experimental dynamic relationships: Recognizing information such as plant gene mutation and overexpression throughout the entire experimental process to establish the dynamic relationship between genes and traits;

[0399] (5) Context-aware reasoning: It can understand the interactive phenotypes of multiple factors such as gene names and crop growth environment in the text, such as discovering implicit relationships such as "Gene X affects yield through pathway Y under low nitrogen conditions."

[0400] Therefore, the large language model 130 of some embodiments of the present invention can perform targeted processing on crop breeding-related data in the literature, so that the extracted triples have richer information.

[0401] In some embodiments, a triple may include a head entity and a tail entity, and the large language model 130 may be configured to extract a gene name or a variety name from the literature as the head entity. In some embodiments, when the head entity is a gene name, the large language model 130 may be configured to extract breeding history, trait records, or cultivation techniques from the literature as the tail entity. In some embodiments, when the head entity is a variety name, the large language model 130 may be configured to extract agronomic traits from the literature as the tail entity.

[0402] When the "head entity" is defined as a gene or variety, the large language model is prevented from extracting other information. Furthermore, detailed regulations are also implemented for the "tail entity." For example, if the "head entity" is a variety, the "tail entity" is limited to breeding history, trait records, cultivation techniques, and other limited aspects. Alternatively, if the "head entity" is a variety, the "tail entity" is limited to agronomic traits (such as plant height, yield, and flowering period), preventing the large language model from extracting irrelevant information. Therefore, the regulations for head and tail entities improve the accuracy of information extraction.

[0403] Crop breeding-related literature obtained using large language models can contain multimodal data, such as text, genetic data, and experimental data. Existing technologies are unable to dynamically link text descriptions with structured omics data (such as gene expression levels and protein interaction networks), limiting their application in tasks such as gene function annotation and breeding strategy generation. Some embodiments of the present invention use large language models to extract triplets from multimodal data, enabling alignment and fusion of multimodal data.

[0404] Here are examples of multimodal data and the triplets that the large language model extracts from the examples of multimodal data:

[0405] Text data (linguistic information consisting of words, sentences, or paragraphs):

[0406] Longjing 31 is a japonica rice variety with 11 leaves on its main stem, a plant height of 92 cm, an ear length of 15.7 cm, approximately 86 grains per ear, and a 1,000-grain weight of 26.3 g. It exhibits strong cold tolerance. Analysis by the Ministry of Agriculture's Rice and Product Quality Supervision and Inspection Center (Harbin) revealed a rough rice yield of 81.1%-81.2%, a head rice yield of 71.6%-71.8%, a chalky grain rate of 0.0%-2.0%, a chalkiness of 0.0%-0.1%, an amylose content (dry basis) of 16.89%-17.43%, a gel consistency of 70.5-71.0 mm, and an eating quality score of 79-82. Inoculation assessment results indicated leaf blast ratings of 3-5 and neck blast ratings of 1-5. Cold tolerance assessment revealed an empty shell rate of 11.39%-14.10%. In the adapted area, the growth period from seedling emergence to maturity is about 130 days, and the active accumulated temperature needs to be ≥10℃ and about 2350℃.

[0407] ===========;

[0408] Plain Text

[0409] ("Longjing 31", "variety", "japonica rice")

[0410] ("Longjing No. 31", "main stem", "11 leaves")

[0411] ("Longjing No. 31", "Ear Length", "15.7cm")

[0412] ("Longjing No. 31", "Number of grains per ear", "86 grains")

[0413] ("Longjing No. 31", "Thousand-grain weight", "26.3g")

[0414] ("Longjing No. 31", "Roughness Rate", "81.1%~81.2%")

[0415] ("Longjing No. 31", "head rice rate", "71.6%~71.8%")

[0416] ("Longjing No. 31", "Chalky Grain Rate", "0.0%~2.0%")

[0417] ("Longjing No. 31", "Chalkyness", "0.0%~0.1%")

[0418] ("Longjing No. 31", "Amylose content", "16.89%~17.43%")

[0419] ("Longjing No. 31", "viscosity", "70.5~71.0mm")

[0420] ("Longjing No. 31", "Taste Quality", "79-82 points")

[0421] ("Longjing No. 31", "Leaf Blight", "Level 3-5")

[0422] ("Longjing No. 31", "Neck Blight", "Level 1-5")

[0423] ("Longjing No. 31", "Handled Empty Shell Rate", "11.39%~14.10%")

[0424] ("Longjing 31", "Birth Days", "130 days")

[0425] ("Longjing No. 31", "Active Accumulated Temperature", "2350℃")

[0426] ===========;

[0427] For example, the large language model's prompts can explicitly instruct it to first identify the "head entity" and then automatically find the "tail entity" and "relationship" based on semantic understanding. Using the text data above as an example, the prompt could include two parts: first, the large language model automatically identifies "Longjing No. 31" as the head entity in the text based on semantic understanding, and then instructs it to find the related "tail entity" and "relationship."

[0428] The following are examples of data from other modalities and the corresponding extracted triples:

[0429] Genomic data (e.g., DNA, RNA sequence data, gene expression data, mutation information, and other biological information):

[0430] AGIS_Os01g000010.mRNA:

[0431] MSSAAGQDNGDTAGDYIKWMCGAGGRAGGAMANLQRGVGSLVRDIGDPCLNPSPVKGSKMLKPEKWHTCFDNDGKVIGFRKALKFIVLGGVDPTIRAEVWEFLLGCYALSSTSEYRRKLRAVRREKYQILVRQCQSMHPSIGTGELAYAVGSKLMDVRTMSKETHIAEEVSTSQQTSQNTAGSLVEDSDYGPGGAQQSQKRESCSKSAELVGFNVHNDTSLYDSSNFIVSSTEVNNCSKDSQDYNDMGEPRYDTETFDDYPSLPVTNFFSTDGVGSNGVDKNHCSFSVPEDRLRHRDERMHSFQINNNIDLIIESNSCSSDVFRASNSDSAIFHSDAYKQDRWLDDNGYNREVIDSLRISDAPEADFVDGTKSNSVVASKDRVSEWLWTLHRIVVDVVRTDSHLDFYGESRNMARMSDILAVYAWVDPSTGYCQGMSDLLSPFVVLYEDDADAFWCFEMLLRRMRENFQMEGPTGVMKQLQALWKIMEITDVELFEHLSTIGAESLHFAFRMLLVLFRRELSFEESLSMWEMMWAADFNEDVILHLEENCLEPLLVDMRNDLSCEVKEEHRVNSYTRRKSKSRKPHHRNGEMRVACNLGMKPNTRNPLCGLSGATIWARHQQMPHISTNVLAKNGDDDLPIFCVAAILVINRHKIIRETRSIDDAIKMFNDNMLKINVKRCVRMAIKLRKKYIYKLLKGGSE

[0432] ==========;

[0433] Plain Text

[0434] ("AGIS_Os01g000010", "amino acid length", "702 amino acids")

[0435] ==========;

[0436] Experimental data (data collected through experimental means, such as numerical data (e.g., measurement values), image data, etc.):

[0437] During the tillering stage of rice, the varieties Pusa-44 (phosphorus deficiency-sensitive type) and NIL-23 (phosphorus deficiency-tolerant type) were compared under sufficient phosphorus (16 ppm) and no phosphorus (0 ppm) conditions. The transcript expression of the gene AGIS_Os10g033440 was upregulated in the stems, leaves, panicles and flowers, while there was a downregulation in the roots.

[0438] ===========;

[0439] Plain Text

[0440] ("AGIS_Os10g033440", "tissue", "stem", "upregulated")

[0441] ("AGIS_Os10g033440", "tissue", "leaf", "upregulated")

[0442] ("AGIS_Os10g033440", "tissue", "spike", "up-regulated")

[0443] ("AGIS_Os10g033440", "tissue", "flower", "up-regulated")

[0444] ("AGIS_Os10g033440", "tissue", "root", "downregulated")

[0445] ===========;

[0446] Therefore, the large language model of some embodiments of the present invention can realize the extraction of triplets of multi-omics data from multimodal data (such as text data, genomic data, experimental data, etc.), thereby realizing deep fusion between different data forms, enabling researchers to process and analyze complex biological problems in a unified system.

[0447] Therefore, the large language model of some embodiments of the present invention solves the problems of data processing and model adaptability in crop breeding research by combining with knowledge graphs, which helps to improve research efficiency and quality in many aspects including breeding, genetic improvement and climate adaptability research.

[0448] In some embodiments, the large language model 130 can be trained by pre-training the large language model base on a general corpus and a crop breeding corpus to obtain an intermediate large language model.

[0449] As an example, the large language model base can be the Tongyi Qianwen 2.5-7B base model, or other suitable base models. The general corpus can include Chinese and English educational corpora and code mathematics special corpora. The crop breeding corpus can include literature related to crop breeding.

[0450] The crop breeding corpus can be extracted using the following methods. First, a search is conducted in Chinese and English databases based on crop breeding-related terms to obtain documents. Then, the documents are converted into text and standardized in format. Based on statistical analysis of the documents, regular expression rules are designed to filter out low-quality text (such as redundant paragraphs). Next, the similarity of text pairs is calculated, and duplicate content is removed at the sentence level. Finally, the texts are classified by subject, retaining only content related to fields such as agriculture, biology, and chemistry. The texts are also scored to eliminate low-scoring content.

[0451] In order to further improve the accuracy of the large language model in answering crop breeding questions, the intermediate large language model can be subjected to supervised learning based on the crop breeding training set to obtain a trained large language model.

[0452] Since the large language model has been trained with a special crop breeding training set, it can accurately understand and process professional terms and complex problems in the field of crop breeding, enhance the knowledge depth, reasoning ability and generation quality of the large language model, and greatly improve the applicability of the large language model in crop breeding research.

[0453] The training process of large language models requires high-quality domain corpora, but the existing corpora have low coverage and lack of multilingual data, which affects the model training effect. In addition, the corpora in crop breeding-related literature usually have the following characteristics:

[0454] (1) Terminology-intensive: Contains a large number of professional terms specific to the field of crop breeding, such as gene names, phenotypic traits, etc.

[0455] (2) Scarcity of structured data: It is difficult to find datasets on the Internet that can be directly used for large language model training, and obtaining high-quality annotated data is costly and time-consuming;

[0456] (3) Complex knowledge associations: involving interactions between multiple factors, such as genes and traits, environmental conditions and yield performance, etc.

[0457] (4) Long-tail knowledge distribution: Some rare breeding techniques or specific variety information appear very rarely in the corpus, but are crucial for professional applications.

[0458] Therefore, it is possible to optimize the performance of large language models for specialized domains by generating training sets relevant to the crop breeding field (also known as synthetic data). Existing data synthesis methods include those based on knowledge graphs and those based on LLMs. Early knowledge graph-based data synthesis methods primarily relied on templates. These methods generated questions from knowledge graphs using predefined templates. While these methods ensured grammatical correctness, the generated questions were rigid, repetitive, lacked diversity, required extensive manual annotation, and had poor scalability. Subsequently, methods based on RNNs (recurrent neural networks) emerged, but they still faced challenges in ensuring factual consistency and generating diverse text. LLM-based methods, on the other hand, suffer from poor controllability and hallucination issues. Although data synthesis methods have emerged that combine seed sets with LLMs—a set of carefully selected or designed initial input samples, typically containing specific tasks or questions—manually constructing seed sets remains time-consuming and labor-intensive.

[0459] Therefore, some embodiments of the present invention propose to use a question-answer extraction model to generate synthetic data.

[0460] In some embodiments, the system 100 may further include a question-answer extraction model 140 .

[0461] The knowledge graph that the large language model 130 searches may be a first knowledge graph. The question-answer extraction model 140 may be configured to extract a second knowledge graph from a crop breeding corpus. The crop breeding corpus includes literature related to crop breeding. The question-answer extraction model 140 may be configured to generate statements based on entities and relationships in the second knowledge graph. The question-answer extraction model 140 may be configured to use the large language model 130 to predict whether the statement is correct to generate a prediction result, and the prediction result indicates the confidence of the large language model 130 in the statement. The question-answer extraction model 140 may be configured to calculate the difference between the prediction result and the true value of the statement. The question-answer extraction model 140 may be configured to determine the statement as a knowledge blind spot of the large language model 130 in response to the difference being greater than an error threshold. The question-answer extraction model 140 may be configured to extract a subgraph in the second knowledge graph based on the knowledge blind spot to generate a crop breeding question-answer sample training set for training the large language model 130.

[0462] As an example, the question-answer extraction model 140 can be a general model, such as Qwen2.5-72B, which can generate multiple statements based on each edge connecting two entities in the second knowledge graph. The statement can be a rewrite of the correct statement of each edge, and the rewritten statement can be correct or wrong. For example, "Long Liangyou 1212 is being demonstrated and planted at the Lvfeng High-Quality Rice Cooperative in Changting County to further promote the variety" generated by the question-answer extraction model 140 is a correct statement. Then, the large language model 130 to be trained can be used to predict whether these statements are "correct" or "wrong", that is, to predict the probability of the large language model 130 answering "yes" or "no", thereby obtaining the confidence of the large language model 130 for each statement. Then, the difference between the confidence and the true value can be calculated as the understanding error. If the understanding error is greater than the error threshold, it means that the difference between the confidence of the large language model 130 on these knowledge and the actual judgment accuracy is large, which means that these knowledge are the knowledge blind spots of the large language model 130. For example, cross-entropy can be calculated to measure the difference between the distribution of confidence scores predicted by the large language model 130 for multiple statements and the distribution of the actual confidence scores for these statements. This difference represents the error in the large language model 130's understanding of these statements. Finally, the question-answer extraction model 140 can extract subgraphs from the second knowledge graph based on knowledge blind spots to generate a training set of crop breeding question-answer samples.

[0463] As an example, a subgraph can be extracted in the following manner. For example, a k-hop subgraph extraction method can be used to extract a subgraph from the second knowledge graph with a certain node as the center and a predefined number of jumps, wherein the composition and scale of the subgraph are adjusted in combination with knowledge blind spots. For example, the number of jumps can be controlled to ensure that the subgraph covers a predefined jump range. Alternatively, the total length of all entity and relationship descriptions in the subgraph can be calculated, and an upper limit can be set to maintain a balanced data distribution. Additionally or alternatively, edges with higher understanding errors can be preferentially selected when expanding the subgraph, thereby increasing potential information gain.

[0464] Additionally, the question-answer extraction model 140 can use different prompt templates to control the style of generated questions and answers, including basic atomic knowledge (such as a single factual statement), complex aggregated knowledge (involving comprehensive analysis of multiple factors), and multi-hop knowledge (requiring reasoning across multiple knowledge nodes).

[0465] The question-answer extraction model of some embodiments of the present invention generates a crop breeding question-answer sample training set by identifying knowledge blind spots and prioritizing the generation of high-value question-answer pairs with long-tail knowledge, thereby addressing the shortcomings of existing methods in terms of coverage, knowledge representation depth, and data diversity, and improving the quality and pertinence of synthetic data.

[0466] Taking rice as an example, after supervised learning of a large language model based on a training set of crop breeding question-answering samples obtained through the above method, the model was tested on multiple common benchmark datasets (CMMLU (Chinese Massive Multitask Language Understanding), GSM8K (Grade School Math 8K), BBH (Big Bench Hard), MMLU (Massive Multitask Language Understanding)), and the Rice Question Answering benchmark. Multiple-choice questions were measured using accuracy, while fill-in-the-blank questions were evaluated using perplexity. Ablation experiments demonstrated that the large language model avoided catastrophic forgetting on the Rice Question Answering test set, significantly improving cross-domain versatility and performance on rice tasks. The BBH metric was 23.6% higher than the baseline.

[0467] To validate the performance of the large language model used in some embodiments of the present invention in answering crop breeding questions, we designed a crop breeding test set at various levels, including basic (e.g., literature retrieval), intermediate (e.g., experimental data analysis), and advanced (e.g., multi-omics integrated reasoning). These test sets cover research areas such as gene function, transcriptomics, proteomics, traditional breeding, and molecular breeding, simulating real-world research scenarios. The following are examples of question answering at different levels, using rice as an example:

[0468] Basic Q&A example:

[0469] ===========;

[0470] Q: How does the disease resistance of Nanjing 46 compare to its female parent?

[0471] A(zh_cn): The disease resistance of Nanjing 46 is similar to that of its parent variety, Wuxiangjing 14, both of which are resistant to leaf stripe disease.

[0472] A(en): The disease resistance of Nanjing 46 is similar to that of its maternal parent Wuxiangjing 14, both possessing the characteristic of resistance to stripe blight.

[0473] ==========;

[0474] Intermediate Q&A Example:

[0475] ==========;

[0476] Q: Pib-mediated defense mechanisms protect rice against Magnaportheoryzae. What role do jasmonic acid-related genes play in this response?

[0477] A: Jasmonic acid-related genes play a crucial role in the Pib-mediated defense mechanisms that protect rice against Magnaporthe oryzae byenhancing the plant's resistance through the jasmonic acid signaling pathway.

[0478] ==========;

[0479] Advanced Q&A Example:

[0480] ==========;

[0481] Q: What were the main proteomic changes observed in rice leaves afterthe rice straw return treatment, and how did these changes affect thephotosynthetic and metabolic processes?

[0482] A: The main proteomic changes observed in rice leaves after the ricestraw return (SR) treatment included increased expression of proteins involved in photosynthesis and metabolic processes. These changes positively affected the photosynthetic efficiency and metabolic activities in the riceleaves.

[0483] ===========;

[0484] Based on the crop breeding test set designed above, the large language model of some embodiments of the present invention has the following experimental results:

[0485] (1) As evaluated by 326 agricultural experts, the large language model of some embodiments of the present invention scored 80.61 on high-level tasks, significantly higher than general models (e.g., DeepSeek-R1 scored 52.37) and surpassed the agricultural undergraduate baseline by 40%;

[0486] (2) Compared with traditional search engines, the response time of the large language model of some embodiments of the present invention is shortened by 90%, the accuracy of answering advanced questions is increased to 88.14%, and the retrieval efficiency is improved;

[0487] (3) After combining with the knowledge graph for crop breeding, the large language model of some embodiments of the present invention achieved an accuracy of 92.3% in gene function prediction tasks, an improvement of 27% over the unenhanced version;

[0488] (4) The large language model of some embodiments of the present invention has a gene identifier parsing accuracy exceeding 95%, and the false hallucination rate is reduced to 12% (compared to 35% for general models);

[0489] (5) In addition to being able to answer crop breeding questions, the large language models of some embodiments of the present invention retain the ability to perform general language understanding, reasoning, and mathematical tasks, achieving accuracy scores comparable to general models in a variety of general knowledge tasks.

[0490] The following are comparison results of the answers given by the large language model of some embodiments of the present invention and the existing model to the same crop breeding question.

[0491] (1) Existing models:

[0492] ===========;

[0493] query: How was Shanyou 63 developed?

[0494] Shanyou 63 is a hybrid rice combination bred by the Sanming Agricultural Science Institute in Fujian Province using Zhenshan 97A and the restorer line Minghui 63.

[0495] ===========;

[0496] (2) Large language models of some embodiments of the present invention:

[0497] ===========;

[0498] <khi>:

[0499] How was Shanyou 63 cultivated?

[0500] <seed11m-20250110 / > :

[0501] The development process of Shanyou 63

[0502] First, 29A is a sterile line of a mid-late indica three-line hybrid rice, primarily used to develop the three-line hybrid rice Yongyou 12. 29A was provided by Researcher Hu Fengcai of the Plant Protection Research and Development Center of the Zhejiang Academy of Agricultural Sciences and was used to extract donor genomic DNA.

[0503] However, 29A not only served as the parent of 25355A, but also provided an important genetic foundation for the development of Shanyou 63. 25355A is a three-line sterile line of medium-grain indica rice, whose genetic foundation is 29A, further contributing to the development of Shanyou 63.

[0504] Shanyou 63, a three-line hybrid rice selected by Mr. Hu Yuqi, has become one of the most widely planted hybrid rice varieties in Chinese history. Mr. Hu selected a highly self-fertile individual from 25355A to serve as the parent, ultimately creating Shanyou 63. Shanyou 63 has not only been widely cultivated in Jiangsu, but its superior traits have also been enhanced through molecular marker selection using modern biotechnology, further enhancing its quality and yield.

[0505] Furthermore, 26151A, as a restorer line for Shanyou 63, provides important genetic support. Shanyou 63, as one of the main parents of C815S and its hybrid combinations, has contributed significant yield advantages.

[0506] The successful development of Shanyou 63 owes its success not only to its excellent genetic foundation but also to the meticulous selection efforts of scientists and the application of modern biotechnology. Mr. Hu Yuqi's contributions were particularly outstanding. By selecting the ideal individual from 25355A and ultimately formulating Shanyou 63, he made a significant contribution to the high and stable yields of rice in China.

[0507] ===========;

[0508] It can be seen from the above answers that the large language model of some embodiments of the present invention can generate more detailed, complete and professional answers.

[0509] It should be understood that although rice is used as an example of crops in the above description, the scope of the present invention is not limited thereto. For example, crops may also include wheat, potatoes, and the like.

[0510] Figure 3 3 is a schematic diagram of a processing flow of a large language model generating an answer based on user input according to some embodiments of the present invention.

[0511] At a first block 310, user input is received. In some embodiments, the large language model may receive a first user input, which may include a first question.

[0512] At the first determination block 320 , it is determined whether the question is related to the field of crop breeding. In some embodiments, the nature of the first question in the first user input may be determined.

[0513] When it is determined that the first question is related to the field of crop breeding, a search is performed in the knowledge graph at the second block 330. In some embodiments, the large language model may use predefined operators to search in the knowledge graph to determine a knowledge graph search result.

[0514] At the second judgment block 340 , it is determined whether the credibility meets the requirements. In some embodiments, the large language model can determine whether the credibility of the knowledge graph retrieval result meets the requirements.

[0515] When it is determined that the credibility of the knowledge graph search result meets the requirements, the knowledge graph search result is returned at the third block 350. In some embodiments, the large language model may return the knowledge graph search result as the final search result.

[0516] Then, at the fourth block 360, the large language model generates an answer. In some embodiments, the large language model may generate an answer based on the first question and the final search results. The above process may be combined with the above Figure 1 The relevant processes described correspond to the following.

[0517] In some embodiments, when it is determined that the credibility of the knowledge graph retrieval result does not meet the requirements (for example, when the judgment result at the second judgment box 340 is "no"), at the fifth box 370, an additional search is performed and the additional search results are returned. In some embodiments, the large language model may perform an additional search on the knowledge graph to determine the additional search results, and return the additional search results as the final search results. Then, the large language model may generate an answer based on the first question and the final search results. The above process can be combined with the above Figure 1 The relevant processes described correspond to the following.

[0518] During the interaction process, if the user inputs a question related to crop breeding, the large language model needs to combine it with a knowledge graph search to generate an answer. This increases processing resource utilization and extends response time, thus affecting the user experience. Therefore, a balance must be struck between obtaining an ideal answer and a good user experience.

[0519] In some embodiments, the large language model can be configured to receive a second input from the user, comprising a second question. The large language model can be configured to determine the relevance of the second question to the field of crop breeding. If the second question is not relevant to the field of crop breeding (e.g., if the determination result at first determination block 320 is "no"), processing can proceed directly to block 360. In some embodiments, the large language model can be configured to generate an answer to the second question based on the second question, without requiring a search within the knowledge graph. For example, if the user inputs a topic related to weather or greetings, the large language model can directly provide a response.

[0520] Whether the question in the user input is related to the field of crop breeding can be judged in the following ways.

[0521] In some embodiments, the large language model can be configured to divide questions in user input into semantic slots. The large language model can extract specific concepts or key information from the natural language text input by the user as semantic slots. Semantic slots are structured semantic information that is typically related to the user's intent, making them easier for machines to understand and process. If a semantic slot is not relevant to the field of crop breeding, the question in the user input can be determined to be unrelated to the field of crop breeding.

[0522] In some embodiments, the large language model can be configured to approximately match the dense features of the question in the user input with the entity word list in the large language model's knowledge base. Dense features refer to the representation of the question (query) mapped to a dense vector through some encoding method, which can efficiently represent complex semantic information. For example, PQ (Product Quantization) and HNSW (Hierarchical Navigable Small World) can be used to accelerate feature comparison. If the approximate matching result is below a predetermined threshold, it can be determined that the question in the user input is not related to the field of crop breeding.

[0523] By filtering user input, the large language model can use its general capabilities to quickly answer questions unrelated to the professional field, making the general performance of the large language model better and saving processing time and resources related to the knowledge graph.

[0524] Therefore, the large language model of some embodiments of the present invention can select the source of the reply content according to the nature of the question. For example, it can come from the knowledge graph, or from the knowledge of the large language model, or it can be a fusion of the knowledge graph and the large language model knowledge, so that it can be dynamically adjusted according to the user's question.

[0525] According to another aspect of the present invention, a method for crop breeding question answering is provided.

[0526] Figure 4 is a flow chart of a first method for answering questions about crop breeding according to some embodiments of the present invention. The first method may be Figure 1 The large language model 130 in the is called upon computing resources 110 for execution.

[0527] The first method may include step S101: receiving input from a user, where the input includes a crop breeding problem.

[0528] The first method may include step S102: using a predefined operator to search in a knowledge graph based on a crop breeding problem to determine a knowledge graph retrieval result, the knowledge graph includes a crop breeding data set, and the predefined operator includes an operation of formal constraints required for retrieving the knowledge graph.

[0529] The first method may include step S103: determining the credibility of the knowledge graph retrieval result.

[0530] The first method may include step S104: selectively performing additional searches based on the confidence level to determine a final search result.

[0531] The first method may include step S105: generating an answer to the crop breeding question based on the crop breeding question and the final search result.

[0532] In some embodiments, a crop breeding dataset may include multi-omics data.

[0533] In some embodiments, the crop breeding dataset may be organized in a knowledge graph in the form of triples, where the triples include entities and relationships, and the entities of the triples of multi-omics data include at least one of the following: genes, transcriptomes, proteins, events, gene annotations, traits, and environmental factors.

[0534] Figure 5 is a flow chart of a first process associated with a first method for crop breeding question answering according to some embodiments of the present invention. The first process may be Figure 1 The large language model 130 in the call computing resources 110 is executed and can be Figure 4 The specific implementation of step S102 in the first method in FIG. 1 is described above, but the scope of the present invention is not limited thereto.

[0535] The first process may include step S1021: decomposing the crop breeding problem into multiple sub-steps.

[0536] The first process may include step S1022: performing an operation of at least one operator in each of the multiple sub-steps to obtain an intermediate search result of each sub-step.

[0537] The first process may include step S1023: determining a knowledge graph retrieval result based on an intermediate retrieval result of a last sub-step among multiple sub-steps.

[0538] In some embodiments, the predefined operators may include at least one of knowledge graph query, arithmetic operation on data, logical operation on data, and result output.

[0539] In some embodiments, the predefined operators can be determined in the following manner: for each evaluation problem in the crop breeding test set, the evaluation problem is decomposed into multiple sub-steps using a large language model, and each sub-step includes an operator; it is determined whether the operator list of the large language model covers the operators of the decomposed sub-steps of all evaluation problems, and the operator list includes predefined operators; and the operator list is selectively updated according to the result of the determination.

[0540] In some embodiments, the predefined operators may be determined by: for each evaluation question in the crop breeding test set, decomposing the evaluation question using a large language model; for evaluation questions that fail to be decomposed, selectively adjusting the expression of operators in the operator list.

[0541] Figure 6 is a flow chart of a second process associated with the first method for crop breeding question answering according to some embodiments of the present invention. The second process may be performed by Figure 1 The large language model 130 in the call computing resources 110 is executed and can be Figure 4 The specific implementation of step S104 in the first method in FIG. 1 is described above, but the scope of the present invention is not limited thereto.

[0542] The second process may include step S1041: determining whether the credibility meets the requirement.

[0543] The second process may include step S1042: in response to the credibility meeting the requirement, returning the knowledge graph retrieval result as the final retrieval result.

[0544] The second process may include step S1043: in response to the credibility not meeting the requirement, performing an additional search on the knowledge graph to determine a final search result.

[0545] In some embodiments, step S1043 may include the following steps: decomposing the crop breeding problem into a multi-level vocabulary, the multi-level vocabulary including first-level vocabulary and second-level vocabulary, the second-level vocabulary including the superordinate concepts of the first-level vocabulary; using the multi-level vocabulary to search in the knowledge graph to determine additional search results; and returning the additional search results as the final search results.

[0546] Figure 7 is a flow chart of a third process associated with the first method for crop breeding question answering according to some embodiments of the present invention. The third process may be performed by Figure 1 The large language model 130 in the call computing resource 110 is executed and may be an additional step of the first process or the second process associated with the first method for crop breeding question answering, but the scope of the present invention is not limited thereto.

[0547] The third process may include step S1024: determining the complexity of the crop breeding problem.

[0548] The third process may include step S1025: in response to the complexity exceeding the complexity threshold, retrieving a predetermined number of relationships associated with the crop breeding problem in the knowledge graph. The predetermined number is based on the length of the longest context of the large language model and the length of the text block mapped to the crop breeding dataset. The predetermined number of relationships includes a predetermined number of relationships with the highest priority, wherein the priorities of the relationships in the knowledge graph are ranked based on the weights of the relationships determined when the knowledge graph was established, and the weights of the relationships reflect the closeness between two entity nodes connected by the relationship in the knowledge graph.

[0549] Figure 8 is a flow chart of a second method for crop breeding question answering according to some embodiments of the present invention. The second method may be Figure 1 The large language model 130 in the is called upon computing resources 110 for execution.

[0550] The second method may include step S201: recording historical messages of a conversation between a user and a large language model, where the historical messages include questions raised by the user.

[0551] The second method may include step S202: detecting the relevance of the question raised by the user with multiple aspects of the field of crop breeding.

[0552] The second method may include step S203: in response to detecting that a threshold number of questions are related to a first aspect among the multiple aspects, increasing the weight of the text block associated with the first aspect in the final retrieval result.

[0553] In some embodiments, the second method can be independent of the above combination Figure 4 In some embodiments, the second method may be combined with the above method. Figure 4 The first method described is performed.

[0554] Figure 9 is a flow chart of a third method for crop breeding question answering according to some embodiments of the present invention. The third method may be Figure 1 The large language model 130 in the is called upon computing resources 110 for execution.

[0555] The third method may include step S301: receiving a second input from a user, where the second input includes a second question.

[0556] The third method may include step S302: determining the relevance of the second question to the field of crop breeding.

[0557] The third method may include step S303: in response to the second question being irrelevant to the field of crop breeding, generating an answer to the second question based on the second question.

[0558] In some embodiments, the large language model is configured to perform the following steps: dividing the second question into semantic slots, and if the semantic slots are not related to the field of crop breeding, determining that the second question is not related to the field of crop breeding.

[0559] In some embodiments, the large language model is configured to perform the following steps: approximately match the dense features of the second question with the entity word list of the knowledge base of the large language model, and if the result of the approximate match is lower than a predetermined threshold, determine that the second question is not related to the field of crop breeding.

[0560] In some embodiments, the third method can be used independently of the above combination Figure 4 In some embodiments, the third method may be combined with the above method. Figure 4 The first method described is performed.

[0561] In some embodiments, the third method can be used independently of the above combination Figure 8 In some embodiments, the third method may be combined with the above method. Figure 8 The second method described is performed.

[0562] Figure 10 is a flow chart of a fourth method for crop breeding question answering according to some embodiments of the present invention. The fourth method may be Figure 1 The large language model 130 in the is called upon computing resources 110 for execution.

[0563] The fourth method may include step S401: obtaining literature related to crop breeding.

[0564] The fourth method may include step S402: extracting triples of a crop breeding dataset from literature, where the triples include a head entity and a tail entity.

[0565] Step S402 may include step S4021: extracting gene names or variety names from the literature as head entities.

[0566] Step S402 may include step S4022: when the head entity is a gene name, extracting breeding history, trait records, or cultivation techniques from literature as a tail entity.

[0567] Step S402 may include step S4023: when the head entity is a variety name, extracting agronomic traits from the literature as a tail entity.

[0568] The fourth method may include step S403: constructing a knowledge graph based on triples of the crop breeding dataset, wherein the triples are mapped to text blocks in the literature.

[0569] In some embodiments, the fourth method can be used independently of the above combination Figure 4 In some embodiments, the fourth method may be combined with the above method. Figure 4 The first method described is performed.

[0570] In some embodiments, the fourth method can be used independently of the above combination Figure 8 In some embodiments, the fourth method may be combined with the above method. Figure 8 The second method described is performed.

[0571] In some embodiments, the fourth method can be used independently of the above combination Figure 9 In some embodiments, the fourth method may be combined with the above Figure 9 The third method described is performed.

[0572] Figure 11 is a flowchart of a fifth method for generating a training dataset according to some embodiments of the present invention.

[0573] In some embodiments, the fifth method can be combined with the above Figure 4 For example, the fifth method can be performed by Figure 1 The question-answer extraction model 140 in the embodiment calls the computing resource 110 to execute. In some embodiments, the fifth method can be independent of the above combination Figure 4 The first method described is performed.

[0574] The fifth method may include step S501: extracting a second knowledge graph from a crop breeding corpus, where the crop breeding corpus includes documents related to crop breeding.

[0575] The fifth method may include step S502: generating statements based on entities and relationships in the second knowledge graph.

[0576] The fifth method may include step S503: using the large language model to predict whether the statement is correct to generate a prediction result, where the prediction result indicates the confidence of the large language model in the statement.

[0577] The fifth method may include step S504: calculating the difference between the predicted result and the true value of the statement.

[0578] The fifth method may include step S505: in response to the difference being greater than the error threshold, determining the statement as a knowledge blind spot of the large language model.

[0579] The fifth method may include step S506: extracting a subgraph in the second knowledge graph based on the knowledge blind spot to generate a crop breeding question and answer sample training set for training the large language model.

[0580] According to another aspect of the present invention, a computer-readable storage medium is provided.

[0581] Figure 12 is a block diagram of a computer-readable storage medium 1200 according to some embodiments of the present invention.

[0582] The computer readable storage medium 1200 stores a computer program 1250. When the computer program 1250 is executed by the processor, the computer program 1250 realizes the above combination Figure 4-11 The steps of each method or process are described.

[0583] According to another aspect of the present invention, a computer program product is provided.

[0584] Figure 13 is a block diagram of a computer program product 1300 according to some embodiments of the present invention.

[0585] The computer program product 1300 may include a computer program 1250. When the computer program 1250 is executed by a processor, the computer program 1250 implements the above-mentioned Figure 4-11 Each step of the method or process is described.

[0586] The embodiments of the present invention have been described with reference to the accompanying drawings, which are intended to be illustrative rather than restrictive.< / khi>

Claims

1. A system for crop breeding question-answering, characterized in that: include: Computing resources; a storage unit configured to store a knowledge graph, the knowledge graph including a crop breeding dataset, wherein the crop breeding dataset includes multi-omics data, the crop breeding dataset is organized in the knowledge graph in the form of triples, the triples include entities and relationships, and the entities of the triples of the multi-omics data include the following: genes, transcriptomes, proteins, events, gene annotations, traits, and environmental factors; and A large language model, wherein the large language model is configured to call the computing resource to: receiving input from a user, the input comprising a crop breeding question; Searching the knowledge graph based on the crop breeding problem using a predefined operator to determine a knowledge graph search result, wherein searching the knowledge graph based on the crop breeding problem using a predefined operator includes: Decomposing the crop breeding problem into multiple sub-steps; Executing at least one operator in each of the plurality of sub-steps to obtain an intermediate search result of each sub-step; and Determine the knowledge graph retrieval result based on the intermediate retrieval result of the last sub-step in the multiple sub-steps, The predefined operator includes an operation of formal constraints required to retrieve the knowledge graph, and the predefined operator includes at least one of a knowledge graph query, an arithmetic operation on data, a logical operation on data, and a result output; Determining the credibility of the knowledge graph retrieval results; Based on the confidence level, selectively performing additional searches to determine final search results; and An answer to the crop breeding question is generated based on the crop breeding question and the final search results.

2. The system according to claim 1, wherein The predefined operator is determined in the following manner: For each evaluation problem in the crop breeding test set, decompose the evaluation problem into multiple sub-steps using the large language model, each sub-step including an operator; determine whether an operator list of the large language model covers all operators of the decomposed sub-steps of the evaluation problem, the operator list including the predefined operators; and selectively update the operator list based on the determination result; or For each evaluation question in the crop breeding test set, the large language model is used to decompose the evaluation question; for the evaluation question that fails to be decomposed, the expression of the operators in the operator list is selectively adjusted.

3. The system according to claim 1, wherein: The large language model is configured to invoke the computing resources to: Determining whether the credibility meets the requirements; In response to the credibility meeting the requirement, returning the knowledge graph retrieval result as the final retrieval result; In response to the credibility not meeting the requirement, performing an additional search on the knowledge graph to determine the final search result.

4. The system according to claim 3, wherein: In response to the credibility not meeting the requirement, the large language model is configured to call the computing resource to: Decomposing the crop breeding problem into a multi-level vocabulary, wherein the multi-level vocabulary includes a first-level vocabulary and a second-level vocabulary, wherein the second-level vocabulary includes hypernyms of the first-level vocabulary; Searching the knowledge graph using the multi-level vocabulary to determine additional search results; as well as The additional search result is returned as the final search result.

5. The system according to claim 1, wherein The large language model is configured to invoke the computing resources to: Determining the complexity of the plant breeding problem; and In response to the complexity exceeding a complexity threshold, a predetermined number of relationships associated with the crop breeding problem are retrieved in the knowledge graph, the predetermined number being based on a length of a longest context of the large language model and a length of a text block mapped to the crop breeding dataset.

6. The system according to claim 5, wherein: The predetermined number of relationships includes the predetermined number of relationships with the highest priority, wherein the priority of the relationships in the knowledge graph is sorted based on the weight of the relationships determined when the knowledge graph is established, and the weight of the relationships reflects the closeness between two entity nodes connected by the relationships in the knowledge graph.

7. The system according to claim 1, wherein: The input is a first input, and the large language model is configured to call the computing resource to: receiving a second input from the user, the second input comprising a second question; Determine the relevance of the second question to the field of crop breeding; as well as In response to the second question being unrelated to the field of crop breeding, an answer to the second question is generated based on the second question.

8. The system according to claim 7, wherein: The large language model is configured to invoke the computing resources to: dividing the second question into semantic slots, and if the semantic slots are not related to the field of crop breeding, determining that the second question is not related to the field of crop breeding; or Approximately matching the dense features of the second question with the entity word list of the knowledge base of the large language model, if the result of the approximate matching is lower than a predetermined threshold, determining that the second question is not related to the field of crop breeding.

9. The system according to claim 1, wherein: The large language model is configured to: Recording historical messages of conversations between the user and the large language model, wherein the historical messages include questions raised by the user; detecting the relevance of the question raised by the user to multiple aspects of the field of crop breeding; as well as In response to detecting that a threshold number of questions are related to a first aspect among the plurality of aspects, a weight of a text block associated with the first aspect in the final search result is increased.

10. The system according to claim 1, wherein: The large language model is configured to: Access to literature related to crop breeding; extracting triplets of the crop breeding dataset from the literature; and The knowledge graph is constructed based on triplets of the crop breeding dataset, wherein the triplets are mapped to text blocks in the document.

11. The system according to claim 10, wherein: The triple includes a head entity and a tail entity, and the large language model is configured to: Extracting gene names or variety names from the literature as head entities; When the head entity is a gene name, breeding history, trait records or cultivation techniques are extracted from the literature as the tail entity; When the head entity is a variety name, agronomic traits are extracted from the literature as the tail entity.

12. The system according to claim 1, wherein The knowledge graph is a first knowledge graph, and the system further includes a question-answer extraction model, wherein the question-answer extraction model is configured to: extracting a second knowledge graph from a crop breeding corpus, wherein the crop breeding corpus includes documents related to crop breeding; generating statements based on entities and relationships in the second knowledge graph; Predicting whether the statement is correct using the large language model to generate a prediction result, wherein the prediction result indicates a confidence level of the large language model for the statement; calculating the difference between the predicted result and the true value of the statement; In response to the difference being greater than an error threshold, determining the statement as a knowledge blind spot of the large language model; as well as Based on the knowledge blind spots, a subgraph is extracted from the second knowledge graph to generate a crop breeding question-answering sample training set for training the large language model.

13. A method for crop breeding question-answering, characterized in that: include: The large language model calls computing resources to perform the following steps: S101: receiving input from a user, wherein the input includes a crop breeding problem; S102: Using a predefined operator, searching the knowledge graph based on the crop breeding problem to determine the knowledge graph search result, the step S102 includes: S1021: Decomposing the crop breeding problem into multiple sub-steps; S1022: Execute at least one operator in each of the multiple sub-steps to obtain an intermediate search result of each sub-step; and S1023: Determine the knowledge graph search result based on the intermediate search result of the last sub-step in the multiple sub-steps, The knowledge graph includes a crop breeding dataset, the crop breeding dataset includes multi-omics data, the crop breeding dataset is organized in the knowledge graph in the form of triples, the triples include entities and relationships, the entities of the triples of the multi-omics data include the following: genes, transcriptomes, proteins, events, gene annotations, traits and environmental factors, the predefined operators include operations of formal constraints required to retrieve the knowledge graph, and the predefined operators include at least one of knowledge graph query, arithmetic operation on data, logical operation on data, and result output; S103: Determine the credibility of the knowledge graph retrieval result; S104: Based on the confidence level, selectively perform additional search to determine a final search result; and S105: Generate an answer to the crop breeding question based on the crop breeding question and the final search result.

14. The method according to claim 13, wherein The predefined operator is determined in the following manner: For each evaluation problem in the crop breeding test set, decompose the evaluation problem into multiple sub-steps using the large language model, each sub-step including an operator; determine whether an operator list of the large language model covers all operators of the decomposed sub-steps of the evaluation problem, the operator list including the predefined operators; and selectively update the operator list based on the determination result; or For each evaluation question in the crop breeding test set, the large language model is used to decompose the evaluation question; for the evaluation question that fails to be decomposed, the expression of the operators in the operator list is selectively adjusted.

15. The method according to claim 13, wherein The step S102 includes: S1024: Determine the complexity of the crop breeding problem; and S1025: In response to the complexity exceeding a complexity threshold, retrieving a predetermined number of relationships associated with the crop breeding problem in the knowledge graph, the predetermined number being based on the length of the longest context of the large language model and the length of the text block mapped to the crop breeding dataset, the predetermined number of relationships including the predetermined number of relationships with the highest priority, wherein the priorities of the relationships in the knowledge graph are sorted based on the weights of the relationships determined when the knowledge graph is established, and the weights of the relationships reflect the closeness between two entity nodes connected by the relationships in the knowledge graph.

16. The method according to claim 13, wherein The large language model calls computing resources to perform the following steps: S201: Recording historical messages of the conversation between the user and the large language model, where the historical messages include questions raised by the user; S202: Detecting the relevance of the question raised by the user with multiple aspects of the crop breeding field; as well as S203: In response to detecting that a threshold number of questions are related to a first aspect among the multiple aspects, increasing the weight of the text block associated with the first aspect in the final search result.

17. The method according to claim 13, wherein The large language model is configured to: S401: Obtain literature related to crop breeding; S402: Extracting triples of the crop breeding dataset from the document, wherein the triples include a head entity and a tail entity, including: S4021: extracting gene names or variety names from the literature as head entities; S4022: When the head entity is a gene name, extracting breeding history, trait records, or cultivation techniques from the literature as a tail entity; S4023: When the head entity is a variety name, extracting agronomic traits from the literature as a tail entity; S403: Constructing the knowledge graph based on the triples of the crop breeding dataset, wherein the triples are mapped to text blocks in the document.

18. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 13 to 17 are implemented.

19. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 13 to 17 are implemented.

Images