A method and device for synthesizing thought chain data for meteorological big language model training

Abstract

This application provides a method and apparatus for synthesizing thought chain data for training a meteorological large language model, relating to the field of artificial intelligence technology. The method includes: obtaining meteorological questions and their corresponding standard answers; deconstructing the standard answers into a meteorological physical causal graph; verifying the natural language description text corresponding to each node based on a configured meteorological law evidence base, obtaining verification results; if any node fails verification, correcting the natural language description text and meteorological state vector corresponding to the node based on the meteorological physical causal graph, obtaining corrected nodes; generating thought chain text based on the nodes that pass verification and the corresponding corrected natural language description text; and determining the thought chain text and meteorological questions as a set of thought chain data for training a meteorological agent. This application can generate high-quality thought chain data for meteorological agents.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and more specifically, to a method and apparatus for synthesizing thought chain data for training meteorological large language models. Background Technology

[0002] With the rapid development of artificial intelligence technology, the application of large language models in the meteorological field is ushering in a historic opportunity. Chain-of-Thought (CoT) data, serving as a reasoning bridge connecting meteorological questions and standard answers, is a core element for fine-tuning the reasoning capabilities of large meteorological models. Existing CoT data available for fine-tuning large model reasoning is typically generated automatically by the large language model based on pre-set prompts, followed by manual verification. CoT data generated in this way cannot be used for training meteorological agents, and its quality is relatively low. Summary of the Invention

[0003] The purpose of this application is to provide a method and apparatus for synthesizing thought chain data for training meteorological big language models, which solves the above-mentioned problems existing in the prior art and can generate high-quality thought chain data for meteorological intelligent agents.

[0004] Firstly, a method for synthesizing thought chain data for training a meteorological big language model is provided, which may include: Obtain the meteorological questions and their corresponding standard answers; The standard answer is deconstructed into a meteorological physics causal graph; the meteorological physics causal graph contains multiple nodes; each node corresponds to a natural language description text and a meteorological state vector; Based on the configured meteorological law evidence database, the natural language description text corresponding to each node is verified to obtain the verification results; If the verification result of any node is that it fails the verification, then the natural language description text and meteorological state vector corresponding to the node are corrected according to the meteorological physical cause-effect graph to obtain the corrected node. Based on the verification results of each node that passed the verification and the corresponding natural language description text of the corrected node, generate the thought chain text; The thought chain text and the meteorological question are identified as a set of thought chain data for training the meteorological agent.

[0005] In an optional implementation, the reverse deconstruction of the standard answer into a meteorological physics causal graph includes: Using a pre-trained deductive generative model, the standard answer is decomposed into multiple causal reasoning steps in causal order, with each causal reasoning step corresponding to a node, and natural language description text for each node is generated. Based on a pre-built meteorological causal model library, semantic parsing, physical quantity extraction, and rule mapping are performed on the natural language description text of each node to obtain the node type, meteorological process model, and meteorological state vector of each node. The node type, meteorological state vector, and natural language description text of each node are used as attributes of each node. Based on the meteorological causal model library, the causal order between each causal reasoning step, and the meteorological process model of each node, directed edges are generated between each node to obtain a meteorological physical causal graph.

[0006] In an optional implementation, the natural language description text corresponding to each node is validated based on the configured meteorological law evidence base, and the validation results are obtained, including: For any node, semantic parsing and physical quantity extraction are performed on the natural language description text of the node to obtain multiple atomic facts and the corresponding physical quantity constraint vectors. Based on the meteorological law evidence database, the natural language description text of each atomic fact, and the physical quantity constraint vector, semantic verification and physical quantity verification are performed on each atomic fact to obtain the verification results of each atomic fact. The node is considered to have passed verification when all atomic facts of the node pass verification.

[0007] In an optional implementation, based on the meteorological law evidence base, the natural language description text of each atomic fact, and the physical quantity constraint vector, semantic verification and physical quantity verification are performed on each atomic fact to obtain the verification results of each atomic fact, including: For any atomic fact, a query representation of the atomic fact is generated based on the natural language description text and physical quantity constraint vector corresponding to the atomic fact; Based on the query representation, semantic channel retrieval and physical constraint channel retrieval are performed in the meteorological law evidence database to obtain multiple candidate evidences and the corresponding text content of the candidate evidences. Calculate the overall matching score and physical quantity difference between the stated atomic facts and each candidate piece of evidence; Based on the natural language description text corresponding to the atomic facts and the text content corresponding to each candidate piece of evidence, determine the semantic relationship probability between the atomic facts and each candidate piece of evidence; The verification result of the atomic fact is determined based on the semantic relationship probability between the atomic fact and each candidate piece of evidence, the comprehensive matching score, and the physical quantity difference.

[0008] In an optional implementation, the physical quantity difference between the atomic facts and each candidate piece of evidence is calculated, including: For any candidate evidence, physical quantities are extracted from the text content corresponding to the candidate evidence to obtain the physical quantity constraint vector of the candidate evidence. For any constraint dimension in the physical quantity constraint vector, the degree of difference between the atomic fact and the candidate evidence under the constraint dimension is calculated based on the values ​​of the physical quantity constraint vector of the atomic fact and the physical quantity constraint vector of the candidate evidence under the constraint dimension. Based on the degree of difference between the atomic fact and the candidate evidence under different constraint dimensions in the physical quantity constraint vector, the degree of difference between the physical quantity of the atomic fact and the candidate evidence is obtained.

[0009] In an optional implementation, the natural language description text and meteorological state vector corresponding to the node are corrected based on the meteorological physical causal graph to obtain the corrected node, including: The atomic facts in the node whose verification result is not passed are taken as the target atomic facts; Based on the semantic relationship probability between the target atomic fact and each candidate piece of evidence, contradictory evidence is screened from multiple candidate pieces of evidence for the target atomic fact. Based on the contradictory evidence, the corrected natural language text of the node is generated, and the meteorological state vector of the node is updated according to the corrected natural language text to obtain the corrected node.

[0010] In an optional implementation, based on the meteorological-physical causal graph, the natural language description text and meteorological state vector corresponding to the node are corrected to obtain the corrected node, which further includes: Based on the meteorological state vectors of the corrected node and the nodes downstream of the corrected node in the meteorological-physical causal graph, the semantic-physical consistency residual is calculated; If the semantic-physical consistency residual is greater than a preset residual threshold, then the natural language description text and meteorological state vector corresponding to the downstream node are corrected according to the meteorological-physical causal graph to obtain the corrected node.

[0011] Secondly, a data synthesis device for training a meteorological large language model is provided, which may include: The acquisition unit is used to acquire meteorological questions and the corresponding standard answers to the meteorological questions; The deconstruction unit is used to reverse-construct the standard answer into a meteorological physics causal graph; the meteorological physics causal graph contains multiple nodes; each node corresponds to a natural language description text and a meteorological state vector; The verification unit is used to verify the natural language description text corresponding to each node according to the configured meteorological law evidence base, and obtain the verification result. The correction unit is used to correct the natural language description text and meteorological state vector corresponding to the node according to the meteorological physical cause-effect graph if the verification result of any node is not passed, so as to obtain the corrected node. The generation unit is used to generate the thought chain text based on the verification results of each node that passed the verification and the corresponding natural language description text of the corrected node. The determining unit is used to determine the thought chain text and the meteorological question as a set of thought chain data for training the meteorological agent.

[0012] Thirdly, an electronic device is provided, which includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; When a processor executes a program stored in memory, it implements any of the steps described in the first aspect above.

[0013] Fourthly, a computer-readable storage medium is provided, wherein a computer program is stored therein, and when executed by a processor, the computer program implements the steps of any of the methods described in the first aspect above.

[0014] This application reverse-engineers the standard answer into a causal graph containing nodes and meteorological state vectors. Each node is automatically validated based on a meteorological law evidence base, with only locally corrected nodes that fail the validation, while retaining the remaining correct nodes. Finally, a thought chain text is generated based on the validated nodes and the corrected nodes. This application effectively improves the quality of thought chain data, ensuring that the generated thought chain data conforms to meteorological law requirements. It meets the needs of meteorological intelligent agents for high-quality reasoning samples, providing high-quality thought chain data for the training of meteorological intelligent agents. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart illustrating a method for synthesizing thought chain data for training a meteorological big language model, provided in an embodiment of this application; Figure 2 A schematic diagram of the structure of a thinking chain data synthesis device for training a meteorological big language model provided in this application embodiment; Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application. Unless otherwise defined, the technical or scientific terms used in this application should have the ordinary meaning understood by those skilled in the art. The words "first," "second," and similar terms used in this application do not indicate any order, quantity, or importance, but are only used to distinguish different components. The words "comprising" or "including," etc., mean that the element or object preceding the word covers the element or object listed after the word and its equivalents, but do not exclude other elements or objects. The words "connected," "coupled," or "connected," etc., are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. "Up," "down," "left," "right," etc., are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0018] The meteorological big data language model training thought chain data synthesis method provided in this application embodiment can be applied to servers and terminals. The server can be a physical server, a server cluster composed of multiple physical servers, or a distributed system. It can also be a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. The terminal can be a user equipment (UE) such as a mobile phone, smartphone, laptop, digital radio receiver, personal digital assistant (PDA), tablet computer (PAD), handheld device, in-vehicle device, wearable device, computing device, or other processing device connected to a wireless modem, mobile station (MS), mobile terminal, etc.

[0019] The preferred embodiments of this application are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit this application. Furthermore, the embodiments and features in the embodiments of this application can be combined with each other without conflict.

[0020] Figure 1 This is a flowchart illustrating a method for synthesizing thought chain data for training a meteorological large language model, provided in an embodiment of this application. Figure 1 As shown, the method may include: Step S110: Obtain the meteorological question and the corresponding standard answer.

[0021] Meteorological questions can be described in natural language and relate to weather mechanisms or meteorological physical processes. Types of meteorological questions include questions about the causes of weather phenomena, deductions of physical processes, changes in meteorological elements, or regional climate characteristics. Meteorological questions are plain text strings of unlimited length and can include information such as place names, meteorological elements, time scales, and phenomenon descriptions. The standard answers to meteorological questions can be reference answers, reference analyses, or manually verified target explanation texts. The standard answers to meteorological questions can include: the causes of the meteorological phenomena in the meteorological question, the physical processes, and the causal chain. The sources of the standard answers to meteorological questions can be original statements in meteorological textbooks or academic literature, standard analyses in meteorological industry standards or operational manuals, explanation texts pre-written by meteorological experts and manually reviewed, high-quality reference answers manually annotated in existing question-and-answer databases, or explanation texts generated by large language models and manually verified for correctness.

[0022] In practice, multiple meteorological questions and their corresponding initial standard answers can be obtained through manual input, semi-automatic filtering, or automatic extraction from public meteorological Q&A datasets, operational forecast texts, expert annotation platforms, or existing knowledge bases. According to the preset text filtering rules, irrelevant annotations, numbers, citation marks, or non-text symbols in each initial standard answer are deleted; if there are formulas or unit symbols in the initial standard answer, the formulas or unit symbols in the initial standard answer are unified into preset standardized expressions according to the preset text conversion rules to obtain each standard answer.

[0023] Step S120: Reverse deconstruct the standard answer into a meteorological physics cause-and-effect diagram.

[0024] The meteorological physics causal graph contains multiple nodes. Each node corresponds to a natural language description text, a node type, and a meteorological state vector; the meteorological state vector can include the meteorological state vector input to the node and the meteorological state vector output by the node.

[0025] In practice, a pre-trained deductive generative model is used to decompose the standard answer into multiple causal reasoning steps according to the causal order. Each causal reasoning step corresponds to a node, and natural language description text for each node is generated. The deductive generative model is a large language model that has been fine-tuned by instructions and lightweightly tuned for the meteorological causal reasoning task. The deductive generative model is trained using meteorological fine-tuning training data. The meteorological fine-tuning training data contains multiple sample pairs manually labeled according to annotation rules. Each sample pair contains a standard answer and a corresponding sequence of causal steps. The annotation rules may include: each step must be semantically independent, meaning that the step expresses a relatively complete causal link; the granularity must be appropriate, meaning that a step does not contain more than two parallel physical processes, nor does it break down a complete physical process too into too many fragments; and the causal order must be clear, meaning that the order of the steps is consistent with... The causal sequence of actual physical events is consistent. Based on a pre-defined prompt template, the standard answer is filled into the corresponding pre-defined positions in the prompt template to construct meteorological reasoning prompts. The deductive generation model, based on the meteorological reasoning prompts, decomposes the standard answer into several causal reasoning steps in causal order. Each causal reasoning step begins with a step sequence, separated by line breaks, without added explanatory text. The step sequence is used to characterize the order of the causal reasoning step in the causal sequence, and can be expressed as "step m," where m is a positive integer. The steps are divided in order, and natural language descriptions of each causal reasoning step are obtained sequentially. When the length of the standard answer exceeds the context window of the deductive generation model, the standard answer can be divided into multiple paragraphs for separate processing, and then the results can be concatenated. Each causal reasoning step corresponds to a node, initially containing only natural language description text. Based on a pre-built meteorological causal model library, semantic parsing, physical quantity extraction, and rule mapping are performed on the natural language description text of each node to obtain the node type, meteorological process model, and meteorological state vector of each node. The meteorological causal model library contains multiple predefined meteorological model sets and multiple objective causal types. Each meteorological model may include: meteorological process name, trigger keyword list, and associated physical quantity list. Meteorological models can include models of meteorological physical processes such as orographic uplift, cooling, saturation, condensation, latent heat release, subsidence over mountains, and adiabatic warming. The meteorological causal model library can be constructed based on the standard physical process definitions in meteorological textbooks and the typical process descriptions in meteorological operational manuals. The order of meteorological models in the meteorological causal model library is pre-set according to the frequency of occurrence of the corresponding meteorological models in the meteorological process, with meteorological models listed earlier having higher priority than those listed later. The meteorological state vector is used to describe the state information of the meteorological system corresponding to the node and is the state representation in the thought chain synthesis process. Objective causal types can include: orographic uplift leading to cooling, cooling leading to saturation, saturation leading to condensation, condensation leading to latent heat release, subsidence over mountains leading to adiabatic warming, water vapor transport leading to humidity changes, and enhanced vertical shear leading to convection organization.

[0026] The node type, meteorological state vector, and natural language description text of each node are used as attributes of each node. Based on the meteorological causal model library, the causal order between each causal reasoning step, and the meteorological process model of each node, directed edges between each node are generated to obtain a meteorological physical causal graph.

[0027] In one embodiment of this application, based on a pre-built meteorological causal model library, semantic parsing, physical quantity extraction, and rule mapping are performed on the natural language description text of each node to obtain the node type, meteorological process model, and meteorological state vector of each node, which may include: For any node, its natural language description text is input into a pre-trained meteorological text classifier, which outputs the node type. The node type can include: initial conditions, physical processes, intermediate states, and conclusions. When the node's natural language description text includes a description of the meteorological background or initial state before the meteorological phenomenon occurs, the node type is initial conditions. Nodes with the initial conditions type often use phrases like "affected by..." or "under the conditions of...". When the node's natural language description text includes a description of a specific physical change or mechanism, the node type is physical processes. Nodes with the physical processes type often include information about uplift. Dynamic verbs such as cooling, condensation, and sinking are used; when the natural language description text of a node contains a description of the result or state change of a certain stage, the node type is intermediate state; the natural language description text of nodes of intermediate state type often uses sentence structures such as "cause…", "make…", or "become…"; when the natural language description text of a node contains the final weather phenomenon or result, the node type is conclusion; the natural language description text of nodes of conclusion type is usually located at the end of the standard answer and often uses sentence structures such as "finally…", "manifest as…", or "form…"; the meteorological text classifier can be a text classification model based on BERT fine-tuning, trained using labeled meteorological text data; The natural language description text of the node is matched with the meteorological causal model library, and the meteorological process model of the node is determined by rule mapping. Specifically, if the natural language description text of the node contains any keyword or phrase from the trigger keyword list of any meteorological model in the meteorological causal model library, then the meteorological model is a candidate meteorological model for the node. The keywords or phrases of the candidate meteorological model contained in the natural language description text of the node are used as target keywords. The candidate meteorological model with the most target keywords is used as the meteorological process model for the node. If there are at least two candidate meteorological models with the same number of target keywords, they are selected according to the priority order in the meteorological causal model library. A meteorological physical quantity extraction model based on named entity recognition is used to identify the numerical and unit combinations in the natural language description text of a node, and the identified numerical and unit combinations are mapped to preset physical quantity fields to obtain the meteorological state vector of the node. The meteorological state vector may include: temperature, pressure layer, relative humidity, specific humidity, altitude or topographic altitude, horizontal wind component, and vertical velocity. The meteorological physical quantity extraction model is fine-tuned on the meteorological text corpus using a pre-trained language model, and the entity types are labeled including temperature, pressure, humidity, altitude, wind speed, and vertical velocity. The extraction results are filled into the corresponding components of the meteorological state vector after uniform unit conversion. The meteorological state vector may include the meteorological state vector input by the node and the meteorological state vector output by the node.

[0028] In another embodiment of this application, the meteorological causal pattern library may include keywords and sentence templates corresponding to different node types. Based on the meteorological causal pattern library, a rule-based method is used to match the natural language description text of each node with the keywords and sentence templates corresponding to different node types in the meteorological causal pattern library to obtain the node type of each node.

[0029] In another embodiment of this application, directed edges between nodes are generated based on a meteorological causal model library, the causal order between each causal inference step, and the meteorological process model of each node, resulting in a meteorological physical causal graph, which may include: Based on the causal order between each causal reasoning step, directed edges are generated between nodes corresponding to adjacent causal reasoning steps; for example, if the step sequence is step1, step2, ..., step... n Then, the initial generation of edges v1→v2, v2→v3,…,v n-1 →v n This indicates the direction of transmission of causal relationships; For any directed edge, its validity is verified according to the meteorological causal model library. If the directed edge passes the validity verification, it is retained; if it fails, the corresponding node is adjusted, and an initial meteorological physical causal graph is generated based on the adjusted nodes and the directed edge. Specifically, a model group consisting of meteorological process models of the two nodes connected by the directed edge is obtained. If the model group is any objective causal type pre-stored in the meteorological causal model library, the directed edge is retained, i.e., the directed edge passes the validity verification; otherwise, the model group is not found in the meteorological causal model library. If any pre-stored objective causal type is found, the directed edge is determined not to satisfy the meteorological causal constraints, meaning the directed edge fails the validity check. For example, if the meteorological process patterns of the two nodes connected by the directed edge are topographic uplift and cooling respectively, the model group is (topographic uplift, cooling), and the model group is the objective causal type of topographic uplift leading to cooling in the meteorological causal model library, then the directed edge is retained. If the meteorological process patterns of the two nodes connected by the directed edge are subsidence warming and condensation respectively, and the model group is not any objective causal type in the meteorological causal model library, then the directed edge is not retained. The initial meteorological physical causal graph is subjected to directed acyclic verification (DAG) on each node and directed edge to obtain the meteorological physical causal graph. Specifically, a topological sorting algorithm is used to traverse all nodes in the initial meteorological physical causal graph. If the topological sort covers all nodes, the initial meteorological physical causal graph satisfies the DAG constraint and is used as the meteorological physical causal graph. If it cannot cover all nodes, it indicates the existence of a cycle in the initial meteorological physical causal graph. The directed edges forming the cycle are reversed or deleted until the initial meteorological physical causal graph satisfies the DAG constraint, thus obtaining the meteorological physical causal graph. The meteorological physical causal graph G=(V,E), where V represents the set of nodes and E represents the set of directed edges. Nodes in the meteorological physical causal graph point from predecessor nodes to successor nodes according to the causal direction, and satisfy the objectivity constraint and DAG constraint of the meteorological causal pattern library. The i-th node in the meteorological physical causal graph... Defined as: ;in, Indicates the node type of the i-th node; This represents the meteorological process model corresponding to the i-th node; The natural language description text of the i-th node; the meteorological state vector of the i-th node. as follows: ;in, Let i be the temperature corresponding to the i-th node. The air pressure layer corresponding to the i-th node. Let be the relative humidity corresponding to the i-th node. The specific humidity corresponding to the i-th node is... The height or terrain height corresponding to the i-th node. and These are the zonal and meridional components of the horizontal wind corresponding to the i-th node, respectively. The vertical velocity corresponding to the i-th node; any directed edge All must satisfy the meteorological causal model library The constraints, namely: ; This represents the j-th node in the meteorological physics causal graph.

[0030] In one embodiment, adjusting the directed edge may include: Split or merge any one or both nodes of the two nodes connected by the directed edge, so that the model group formed by the meteorological process models of the adjacent nodes after adjustment is an objective causal type in the meteorological causal model library; split any one node of the two nodes connected by the directed edge into at least two child nodes, and generate natural language description text and meteorological process model for each child node, such that in the split node sequence, the directed edges between adjacent nodes, including the original node, sequentially satisfy the objective causal type in the meteorological causal model library; for example, if the meteorological process model of the first node of the two nodes connected by the directed edge is orographic uplift, the second node The meteorological process pattern is condensation, but the natural language description text of the first node also contains information on uplift, cooling, and saturation. Therefore, the first node is split into three sub-nodes: topographic uplift, cooling, and saturation, and directed edges are generated. This ensures that the pattern groups (topographic uplift, cooling) and (cooling, saturation) formed by the meteorological process patterns corresponding to the three sub-nodes are the two objective causal types in the meteorological causal model library: topographic uplift leading to cooling and cooling leading to saturation. Furthermore, the pattern group (saturation, condensation) formed by one of the sub-nodes and the meteorological process pattern corresponding to the second node is also the two objective causal types in the meteorological causal model library: cooling leading to saturation and saturation leading to condensation. Node merging is applicable when the meteorological process patterns of two adjacent nodes describe different stages of the same physical process and cannot independently satisfy the objective causal type constraints. During merging, the natural language description texts of the two nodes are concatenated in causal order, the merged meteorological process pattern and meteorological state vector are re-extracted, and the directed edges between the merged node and its upstream and downstream nodes are re-verified to satisfy the constraints of the meteorological causal model library.

[0031] Step S130: Based on the configured meteorological law evidence database, verify the natural language description text corresponding to each node to obtain the verification result.

[0032] The meteorological law evidence database can contain content from meteorological textbooks, industry standards, operational manuals, reanalysis data documentation, and selected academic papers. The meteorological law evidence database includes multiple pieces of evidence, along with the text content, source, and authority level of each piece of evidence. Sources can include meteorological textbooks, industry standards, operational manuals, selected academic papers, and online materials. Each source corresponds to an authority level.

[0033] In practice, for any given node, semantic parsing and physical quantity extraction are performed on the natural language description text of that node to obtain multiple atomic facts and corresponding physical quantity constraint vectors. Each atomic fact expresses only one independent physical judgment and does not simultaneously contain two or more parallel conclusions. A physical judgment is a verifiable minimal meteorological physical proposition jointly expressed by the object of action, action or state change, and physical quantity constraints. Each atomic fact expresses only one independent physical judgment and does not simultaneously contain two or more parallel conclusions. If a text segment contains two or more parallel action, state change, or physical process tags, it is split into multiple atomic facts. Each atomic fact is bound to at least one meteorological physical quantity or physical process tag. Each atomic fact can be mapped back to its own node and its upstream and downstream causal relationships. Based on the meteorological law evidence database, the natural language description text of each atomic fact, and the physical quantity constraint vector, semantic verification and physical quantity verification are performed on each atomic fact to obtain the verification results of each atomic fact. Specifically, for any atomic fact, a query representation of the atomic fact is generated based on the natural language description text and physical quantity constraint vector corresponding to the atomic fact; wherein, the natural language description text corresponding to the atomic fact is input into the semantic encoder to obtain the text semantic vector; the text semantic vector is concatenated or attention fused with the physical quantity constraint vector to obtain the query representation of the atomic fact. Based on the query representation, semantic channel retrieval and physical constraint channel retrieval are performed in the meteorological law evidence database to obtain multiple candidate evidences and their corresponding text content. Specifically, a meteorological physical quantity extraction model based on named entity recognition is used to extract physical quantities from the text content corresponding to each candidate evidence, resulting in physical quantity constraint vectors for each candidate evidence. Cosine similarity matching is performed between the text semantic vectors in the query representation and the text semantic vectors of each piece of evidence in the meteorological law evidence database. Distance metric matching is performed between the physical quantity constraint vectors in the query representation and the physical quantity constraint vectors of each piece of evidence in the meteorological law evidence database. After fusing and sorting the results of the semantic channel retrieval and the physical constraint channel retrieval, the top-ranked evidences are selected as candidate evidences. Based on the query representation of the atomic fact and the text content corresponding to each candidate piece of evidence, calculate the comprehensive matching score and physical quantity difference between the atomic fact and each candidate piece of evidence; Based on the natural language description text corresponding to the atomic fact and the text content corresponding to each candidate piece of evidence, the semantic relationship probability between the atomic fact and each candidate piece of evidence is determined. Specifically, the natural language description text of the atomic fact and the text content of each candidate piece of evidence are combined into an input pair, and this input pair is input into a pre-trained natural language reasoning and discrimination model, which outputs the implication probability, contradiction probability, and neutral probability of the candidate piece of evidence relative to the atomic fact. Based on the implication probability, contradiction probability, and neutral probability of the atomic fact, the semantic relationship probability between the atomic fact and each candidate piece of evidence is generated. The natural language reasoning and discrimination model is obtained by fine-tuning a pre-trained language model, and the training data is three-class labeled data of implication, contradiction, and neutrality in the meteorological field. The output layer is a 3-dimensional softmax, corresponding to the implication probability, contradiction probability, and neutral probability, respectively. The pre-trained language model can be a BERT model. Based on the semantic relationship probability, comprehensive matching score, and physical quantity difference degree of the atomic fact with each candidate piece of evidence, the verification result of the atomic fact is determined. Specifically, when there is at least one candidate piece of evidence and the atomic fact whose semantic relationship probability is not less than the first preset threshold, whose contradiction probability is less than the second preset threshold, and whose physical quantity difference degree is not greater than the third preset threshold, the atomic fact is determined to pass the verification; otherwise, the atomic fact is determined to fail the verification; if it cannot be clearly determined, it is determined to be uncertain. Among them, the first preset threshold can be 0.7, which is the implication probability threshold, used to control the minimum semantic support strength required for the atomic fact to pass the verification; the second preset threshold can be 0.6, which is the contradiction probability threshold, used to control the maximum semantic contradiction degree allowed for the atomic fact to pass the verification; the third preset threshold can be 1.0, which is the physical quantity difference degree threshold, used to control the maximum physical quantity deviation degree allowed for the atomic fact to pass the verification. The node's verification result is considered passed when all atomic facts of the node pass the verification.

[0034] In one embodiment of this application, semantic parsing and physical quantity extraction are performed on the natural language description text of the node to obtain multiple atomic facts and corresponding physical quantity constraint vectors, which may include: Semantic role parsing is performed on the natural language description text of this node. The natural language description text is divided into multiple smallest semantic units according to independent physical judgments. Each smallest semantic unit corresponds to an atomic fact. Semantic role parsing can be performed using a semantic role annotation tool based on a pre-trained language model. By identifying the predicate-argument structure in the natural language description text, each predicate and its core argument are combined as an independent atomic fact. For any given atomic fact, a physical quantity constraint vector is extracted from the corresponding natural language description text. Specifically, a physical quantity extraction model based on named entity recognition is used to identify multiple combinations of numerical values ​​and units in the text. Each combination of numerical values ​​and units is mapped to a preset physical quantity field to obtain multiple initial physical quantity components. Each initial physical quantity component is standardized to obtain a physical quantity constraint vector. The physical quantity constraint vector may include: pressure layer constraints, altitude or terrain altitude constraints, temperature constraints, relative humidity constraints, specific humidity constraints, time scale constraints, regional or terrain identifiers, and process type identifiers. The physical quantity constraint vector is used to describe the constraints that the atomic fact must satisfy to be valid and serves as the basis for evidence retrieval and consistency verification. All atomic facts obtained after splitting the node are stored as the atomic fact set of the node, and the mapping relationship between each atomic fact and the corresponding node, as well as the upstream and downstream nodes of the node, is recorded so that the source of the error can be traced and the specific atomic fact and node can be located when the verification fails.

[0035] In one embodiment of this application, an atomic fact may include: an object of action, an action or state variable, and a condition for its validity. Semantic role resolution can employ a Span-based semantic role annotation model to identify each predicate in the natural language description text and extract the corresponding agent, patient, time, location, and manner argument roles, extracting each predicate and argument structure as an independent atomic fact. Here, the predicate corresponds to an action or state change variable; the agent or patient corresponds to the object of action; and the additional arguments such as time, location, and manner combine to form the condition for its validity. For example, for the natural language description text "Airflow descends after crossing a ridge, with a warming rate of approximately 0.5 to 0.6°C / 100m according to the wet adiabatic lapse rate," the semantic role resolution model identifies two independent physical judgments: one is that the airflow descends after crossing a ridge, and the other is that the warming rate is approximately 0.5 to 0.6°C / 100m according to the wet adiabatic lapse rate. The two judgments have different predicates (descending and warming) and each expresses a complete physical process, thus being split into two atomic facts.

[0036] In another embodiment of this application, the physical quantity extraction model can employ a BERT-based named entity recognition model, and fine-tune the model using meteorological text corpus. The entity types are labeled with temperature values, air pressure values, humidity values, altitude values, wind speed values, time values, region names, and process type keywords. The extraction results are then converted to uniform units and filled into the corresponding components of the physical quantity constraint vector.

[0037] In another embodiment of this application, calculating the comprehensive matching score and physical quantity difference between the atomic fact and each candidate piece of evidence based on the query representation of the atomic fact and the text content corresponding to each candidate piece of evidence may include: For any candidate evidence, a meteorological physical quantity extraction model based on named entity recognition is used to extract physical quantities from the text content corresponding to the candidate evidence, thereby obtaining the physical quantity constraint vector of the candidate evidence. The cosine similarity between the text semantic vector in the query representation of the atomic fact and the text semantic vector of the corresponding text content of each candidate evidence is calculated to obtain the semantic matching component. The first target constraint and the second target constraint are extracted from the physical quantity constraint vector of the candidate evidence and the physical quantity constraint vector of the atomic fact, respectively; wherein, the target constraints may include: barosphere constraints, altitude constraints and process type identifiers; Calculate the difference between the first objective constraint and the second objective constraint; Based on the difference and preset matching rules, a comprehensive matching score is calculated. Specifically, a pressure layer matching component is generated based on whether the pressure layer difference between the pressure layer constraint of the candidate evidence and the pressure layer constraint of the atomic fact exceeds a preset pressure difference threshold. If the pressure layer difference does not exceed the preset pressure difference threshold, the pressure layer matching component is 1; otherwise, the pressure layer matching component is 0. A height matching component is generated based on whether the height difference between the height constraint of the candidate evidence and the height constraint of the atomic fact exceeds a preset height difference threshold. If the height difference does not exceed the preset height difference threshold, the height matching component is 1; otherwise, the height matching component is 0. The matching component is 0; a process type matching component is generated based on whether the process type identifier of the candidate evidence and the process type identifier of the atomic fact are the same; if they are the same, the process type matching component is 1, otherwise, the process type matching component is 0; the semantic matching component, the pressure layer matching component, the altitude matching component, and the process type matching component are multiplied by their respective weight coefficients and then summed to obtain the comprehensive matching score; the preset altitude difference threshold and the preset pressure difference threshold can be dynamically adjusted according to the confidence level of different meteorological tasks, meteorological law evidence bases, and operational fault tolerance requirements; the comprehensive matching score calculation formula is as follows:

[0038] in, This represents the r-th candidate evidence and the k-th atomic fact of the i-th node. The overall matching score; and These represent the weight coefficients of the semantic matching component, the barosphere matching component, the altitude matching component, and the process type matching component, respectively. and It can take values ​​of 0.5, 0.2, 0.2, and 0.1 respectively, and satisfies the following conditions: + + + =1; Each weight coefficient can be dynamically adjusted according to the confidence level of the meteorological law evidence database and the business fault tolerance requirements; Semantic encoding of the natural language description text representing the k-th atomic fact of the i-th node; The semantic encoding of the text content representing the r-th candidate evidence; The physical quantity constraint vectors representing the k-th atomic fact of the i-th node are respectively. The atmospheric pressure layer constraints, altitude constraints, and process type identifiers within the framework; These represent the pressure layer constraint, altitude constraint, and process type identifier constraint in the physical quantity constraint vector of the r-th candidate evidence, respectively. The preset air pressure difference threshold; The preset height difference threshold is used; This is an indicator function that outputs 1 when the condition inside the parentheses is true and 0 when it is false. Indicate and Cosine similarity between them The range of its value is [-1, 1], with [0, 1] being preferred; For any candidate evidence, a meteorological physical quantity extraction model based on named entity recognition is used to extract physical quantities from the text content corresponding to the candidate evidence, thereby obtaining the physical quantity constraint vector of the candidate evidence. The cosine similarity between the text semantic vector in the query representation of the atomic fact and the text semantic vector of the corresponding text content of each candidate evidence is calculated to obtain the semantic matching component. The first target constraint and the second target constraint are extracted from the physical quantity constraint vector of the candidate evidence and the physical quantity constraint vector of the atomic fact, respectively; wherein, the target constraints may include: barosphere constraints, altitude constraints and process type identifiers; For any constraint dimension in the physical quantity constraint vector, calculate the absolute value of the difference between the physical quantity constraint vector of the atomic fact and the physical quantity constraint vector of the candidate evidence under that constraint dimension, and take the ratio of the absolute value of the difference to the preset normalization scale of the constraint dimension as the normalization difference of the constraint dimension. The product of the normalized difference and the preset weight of the constraint dimension is used as the degree of difference between the atomic fact and the candidate evidence under the constraint dimension. Based on the degree of difference between the atomic fact and the candidate evidence under different constraint dimensions, the physical quantity difference is obtained. Specifically, the sum of the degrees of difference between the atomic fact and the candidate evidence under all constraint dimensions is taken as the physical quantity difference. The smaller the value of the physical quantity difference, the more consistent the physical quantity constraints are between the atomic fact and the candidate evidence. The formula for calculating the physical quantity difference is as follows: ; in, This represents the degree of difference in physical quantity between the k-th atomic fact of the i-th node and the r-th candidate evidence; The l-th physical quantity constraint vector represents the k-th atomic fact of the i-th node; This represents the constraint vector of the l-th physical quantity for the r-th candidate evidence; that is... and Let them represent the values ​​of the physical quantity constraint vector of the k-th atomic fact of the i-th node and the physical quantity constraint vector of the r-th candidate evidence under the same constraint dimension l, respectively. This represents the set of constraint dimensions in the physical quantity constraint vector jointly contained by the k-th atomic fact and the r-th candidate evidence of the i-th node; This represents the preset weight of the l-th constraint dimension, with a value range of 0-1; This represents the preset normalization scale for the l-th constraint dimension.

[0039] In another embodiment of this application, the verification result of the atomic facts is obtained by the atomic fact verification function judging whether the atomic facts conform to the meteorological law evidence base based on the semantic relationship probability, comprehensive matching score, and physical quantity difference degree; wherein, the atomic fact verification function is as follows:

[0040] in, This represents the verification result of the k-th atomic fact of the i-th node; This indicates that the atomic fact has passed the verification; This indicates that the atomic fact failed the verification. Indicates the implied probability; This represents the maximum value among the implied probabilities of each candidate piece of evidence relative to the atomic fact. This indicates the first preset threshold. Indicates the probability of a contradiction; This represents the maximum value among the implied probabilities of each candidate piece of evidence relative to the atomic fact. This indicates the second preset threshold. This indicates the third preset threshold. The physical quantity difference between the k-th atomic fact of the i-th node and the r-th candidate evidence is represented; the first preset threshold can be 0.7; the second preset threshold can be 0.6; and the third preset threshold can be 1.0. This indicates uncertainty; neither the PASS nor FAIL condition is met. The verification result of the k-th atomic fact of the i-th node is uncertain.

[0041] Step S140: If the verification result of any node is that it fails the verification, then the natural language description text and meteorological state vector corresponding to the node are corrected according to the meteorological physical cause-effect graph to obtain the corrected node.

[0042] In practice, atomic facts in a node that fail the verification are taken as target atomic facts. Based on the semantic relationship probability between the target atomic fact and each candidate piece of evidence, contradictory evidence is screened from multiple candidate pieces of evidence for the target atomic fact. Specifically, the source and source authority level of each candidate piece of evidence are obtained. Based on the source, source authority level, comprehensive matching score and contradiction probability of each candidate piece of evidence, the target contradictory evidence is screened from multiple candidate pieces of evidence. Among them, when comprehensively ranking all candidate pieces of evidence, they are first grouped by source authority level, and then sorted by comprehensive matching score from high to low within the same authority level. For evidence with similar comprehensive matching scores, the one with the highest contradiction probability is selected as the target contradictory evidence. Based on contradictory evidence, a corrected natural language text for the node is generated. The meteorological state vector of the node is then updated based on this corrected text to obtain the corrected node. While keeping the natural language descriptions and meteorological state vectors of all upstream predecessor nodes of the node unchanged, the node type and meteorological process pattern of the node remain unchanged; only the natural language description text of the node is corrected and regenerated. A set of meteorological physical constraints for the node is dynamically generated based on the meteorological law evidence base, the meteorological causal pattern base, the meteorological process pattern of the node, and the meteorological state vector. The target contradictory evidence, the node's natural language description text, and the set of meteorological physical constraints are input into a pre-trained deductive generation model to obtain the corrected natural language text for the node. A physical quantity extraction model based on named entity recognition is invoked to re-extract the meteorological state vector from the corrected natural language text and update the meteorological state vector of the node in the meteorological physical causal graph. The updated meteorological state vector, the corrected natural language text, and the node type and meteorological process pattern of the node are combined to obtain the corrected node. The set of meteorological physical constraints includes at least: The pressure layer is aligned with the direction of altitude change; the unsaturated sinking warming rate is not less than the saturated warming rate; phase change nodes such as condensation, melting, and evaporation are aligned with temperature and humidity conditions; the input physical quantities of downstream nodes must be deriveable from upstream correction nodes.

[0043] The semantic-physical consistency residual is calculated based on the meteorological state vectors of the corrected node and the nodes downstream of the corrected node in the meteorological-physical causal graph. Specifically, the semantic-physical consistency residual is calculated based on the meteorological state vector output by the corrected node and the meteorological state vector input by the nodes downstream of the corrected node in the meteorological-physical causal graph. The semantic-physical consistency residual is used to quantify the impact of the upstream node's correction on the downstream node's input state: if the residual is small, it indicates that the impact of the upstream node's correction on the downstream node's input state is within an acceptable range, and the downstream node does not need to be adjusted; if the residual is large, it indicates that the upstream node's correction has significantly changed the input conditions on which the downstream node depends, and the downstream node needs to be adjusted accordingly. The formula for calculating the semantic-physical consistency residual is as follows: ;in, This represents the meteorological state vector input to the b-th node downstream of the corrected node in the meteorological-physical causal graph, with the superscript in indicating the input meteorological state vector of the b-th node. This represents the node obtained after the correction of the a-th node. The input meteorological state vector; This represents the node obtained after the correction of the a-th node. The node located in the meteorological causality diagram after correction The semantic-physical consistency residual of the b-th downstream node; if the semantic-physical consistency residual is greater than the preset residual threshold, the natural language description text and meteorological state vector corresponding to the downstream node are corrected according to the meteorological physical causal graph to obtain the corrected node; wherein, the preset residual threshold ranges from 0.2 to 0.5, preferably 0.35, and the preset residual threshold can be dynamically adjusted according to the accuracy requirements of the meteorological task and the confidence of the meteorological law evidence base. When the meteorological task has high requirements for physical consistency, the preset residual threshold can be appropriately reduced; since the graph structure satisfies the directed acyclic feature, the propagation path of local replanning proceeds unidirectionally along the topological order, and the maximum number of propagation rounds does not exceed the total number of nodes; specifically, for all downstream successor nodes that are triggered for replanning, the above correction operation is repeated, that is, the natural language description text of the downstream successor node is regenerated based on the target contradiction evidence corresponding to the atomic facts that have not passed the verification in the downstream successor node, and the meteorological state vector of the downstream successor node is updated to obtain the corrected downstream node.

[0044] In one embodiment of this application, after selecting target contradictory evidence from multiple candidate evidences of the target atomic facts, the method may further include: constructing an error description tuple based on the target atomic facts, target contradictory evidence, error constraint dimension, and physical quantity difference degree; correcting and regenerating the natural language description text of the node based on the target contradictory evidence, error constraint dimension, and physical quantity difference degree in the error description tuple, and updating the meteorological state vector of the node based on the corrected text to obtain the corrected node; the error constraint dimension is the physical quantity dimension or process type that causes the conflict.

[0045] Step S150: Generate the thought chain text based on the natural language description text corresponding to each node that passed the verification and the corresponding corrected node.

[0046] Step S160: Determine the thought chain text and the meteorological question as a set of thought chain data for training the meteorological agent.

[0047] In one embodiment of this application, after all nodes in the meteorological physics causal graph have passed the verification, the node text is causally serialized according to the topological sorting to generate the final thought chain text.

[0048] This application does not simply perform segmented retrieval of meteorological answers, but rather uses a meteorological physical causal graph as a carrier to refine the retrieval verification to the atomic fact layer with physical quantity constraints. When conflicts occur, local reprogramming is only performed on erroneous nodes and necessary successor nodes. This scheme can more effectively suppress hidden physical errors in complex meteorological reasoning data, improving the credibility, traceability, and scalability of thought chain samples.

[0049] This application addresses complex meteorological question-and-answer samples by reverse-engineering standard answers into meteorological physical causal graphs. It then performs atomic fact retrieval and verification on the nodes in the graph based on constraints related to pressure layers, temperature, humidity, and topography. When physical conflicts are detected, a thought-chain data synthesis method is implemented, performing local replanning only on the conflicting nodes and their affected successor nodes. This method is applicable to scenarios involving topographic weather, severe convection, heavy rain, cold waves, typhoons, and other complex weather mechanisms.

[0050] This application provides a method for synthesizing thought chain data based on atomic fact verification and local replanning of meteorological physical causal graphs. This method uses meteorological physical causal graphs as an intermediate carrier, further breaking down nodes into atomic facts constrained by pressure layers, temperature and humidity, topography, and process type. When an atomic fact contradicts authoritative meteorological data, local replanning is performed only on the corresponding node and its necessary successor nodes, thereby reducing the granularity of error localization from the entire thought chain to a single atomic fact, and reducing the regeneration scope from the entire graph to a locally affected subgraph.

[0051] Corresponding to the above method, this application also provides a mind chain data synthesis device for training a meteorological large language model, such as... Figure 2 As shown, the device includes: Unit 210 is used to obtain meteorological questions and their corresponding standard answers. Deconstruction unit 220 is used to reverse-construct the standard answer into a meteorological physics causal graph; the meteorological physics causal graph contains multiple nodes; each node corresponds to a natural language description text and a meteorological state vector; The verification unit 230 is used to verify the natural language description text corresponding to each node according to the configured meteorological law evidence database, and obtain the verification result. The correction unit 240 is used to correct the natural language description text and meteorological state vector corresponding to the node according to the meteorological physical causal graph if the verification result of any node is not passed, so as to obtain the corrected node. The generation unit 250 is used to generate the thought chain text based on the natural language description text corresponding to each node that passes the verification and the corresponding corrected node. Unit 260 is used to identify the thought chain text and meteorological questions as a set of thought chain data for training the meteorological agent.

[0052] The functions of each functional unit of the meteorological big language model training thought chain data synthesis device provided in the above embodiments of this application can be realized through the above methods and steps. Therefore, the specific working process and beneficial effects of each unit in the meteorological big language model training thought chain data synthesis device provided in the embodiments of this application will not be repeated here.

[0053] This application also provides an electronic device, such as... Figure 3 As shown, it includes a processor 310, a communication interface 320, a memory 330, and a communication bus 340, wherein the processor 310, the communication interface 320, and the memory 330 communicate with each other through the communication bus 340.

[0054] Memory 330 is used to store computer programs; When the processor 310 executes the program stored in the memory 330, it performs the following steps: Get meteorological questions and their standard answers; The standard answer is deconstructed into a meteorological physics causal graph; the meteorological physics causal graph contains multiple nodes; each node corresponds to a natural language description text and a meteorological state vector; Based on the configured meteorological law evidence database, the natural language description text corresponding to each node is verified to obtain the verification results; If the verification result of any node is not passed, the natural language description text and meteorological state vector corresponding to the node are corrected according to the meteorological physical cause-effect graph to obtain the corrected node. Based on the verification results of each node that passed the verification and the corresponding natural language description text of the corrected node, generate the thought chain text; The thought chain text and meteorological questions were identified as a set of thought chain data for training the meteorological agent.

[0055] The communication bus mentioned above can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not mean that there is only one bus or one type of bus.

[0056] The communication interface is used for communication between the aforementioned electronic devices and other devices.

[0057] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0058] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0059] The implementation methods and beneficial effects of the various components of the electronic device in the above embodiments for solving the problem can be found in [reference needed]. Figure 1 The steps in the illustrated embodiments are used to implement the electronic device. Therefore, the specific working process and beneficial effects of the electronic device provided in this application will not be repeated here.

[0060] In another embodiment provided in this application, a computer-readable storage medium is also provided, which stores instructions that, when executed on a computer, cause the computer to perform the thought chain data synthesis method for training the meteorological big language model as described in any of the above embodiments.

[0061] In another embodiment provided in this application, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute the thought chain data synthesis method for training the meteorological big language model as described in any of the above embodiments.

[0062] Those skilled in the art will understand that the embodiments in this application can be provided as methods, systems, or computer program products. Therefore, the embodiments in this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the embodiments in this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0063] This application describes embodiments of methods, apparatus (systems), and computer program products according to embodiments of this application with reference to flowchart illustrations and / or block diagrams. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0064] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0065] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0066] Although preferred embodiments have been described in this application, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of this application.

[0067] Obviously, those skilled in the art can make various modifications and variations to the embodiments of this application without departing from the spirit and scope of the embodiments of this application. Therefore, if these modifications and variations to the embodiments of this application fall within the scope of this application and its equivalents, then these modifications and variations are also intended to be included in the embodiments of this application.

Claims

1. A method for synthesizing thought chain data for training a meteorological large language model, characterized in that, The method includes: Obtain the meteorological questions and their corresponding standard answers; The standard answer is deconstructed into a meteorological physics causal graph; the meteorological physics causal graph contains multiple nodes; each node corresponds to a natural language description text and a meteorological state vector; Based on the configured meteorological law evidence database, the natural language description text corresponding to each node is verified to obtain the verification results; If the verification result of any node is that it fails the verification, then the natural language description text and meteorological state vector corresponding to the node are corrected according to the meteorological physical cause-effect graph to obtain the corrected node. Based on the verification results of each node that passed the verification and the corresponding natural language description text of the corrected node, generate the thought chain text; The thought chain text and the meteorological question are identified as a set of thought chain data for training the meteorological agent.

2. The method as described in claim 1, characterized in that, The process of reversing the standard answer into a meteorological physics causal diagram includes: Using a pre-trained deductive generative model, the standard answer is decomposed into multiple causal reasoning steps in causal order, with each causal reasoning step corresponding to a node, and natural language description text for each node is generated. Based on a pre-built meteorological causal model library, semantic parsing, physical quantity extraction, and rule mapping are performed on the natural language description text of each node to obtain the node type, meteorological process model, and meteorological state vector of each node. The node type, meteorological state vector, and natural language description text of each node are used as attributes of each node. Based on the meteorological causal model library, the causal order between each causal reasoning step, and the meteorological process model of each node, directed edges are generated between each node to obtain a meteorological physical causal graph.

3. The method as described in claim 1, characterized in that, Based on the configured meteorological law evidence database, the natural language description text corresponding to each node is validated, and the validation results are obtained, including: For any node, semantic parsing and physical quantity extraction are performed on the natural language description text of the node to obtain multiple atomic facts and the corresponding physical quantity constraint vectors. Based on the meteorological law evidence database, the natural language description text of each atomic fact, and the physical quantity constraint vector, semantic verification and physical quantity verification are performed on each atomic fact to obtain the verification results of each atomic fact. The node is considered to have passed verification when all atomic facts of the node pass verification.

4. The method as described in claim 3, characterized in that, Based on the meteorological law evidence database, the natural language description text of each atomic fact, and the physical quantity constraint vector, semantic verification and physical quantity verification are performed on each atomic fact to obtain the verification results of each atomic fact, including: For any atomic fact, a query representation of the atomic fact is generated based on the natural language description text and physical quantity constraint vector corresponding to the atomic fact; Based on the query representation, semantic channel retrieval and physical constraint channel retrieval are performed in the meteorological law evidence database to obtain multiple candidate evidences and the corresponding text content of the candidate evidences. Calculate the overall matching score and physical quantity difference between the stated atomic facts and each candidate piece of evidence; Based on the natural language description text corresponding to the atomic facts and the text content corresponding to each candidate piece of evidence, determine the semantic relationship probability between the atomic facts and each candidate piece of evidence; The verification result of the atomic fact is determined based on the semantic relationship probability between the atomic fact and each candidate piece of evidence, the comprehensive matching score, and the physical quantity difference.

5. The method as described in claim 4, characterized in that, Calculate the physical quantity difference between the stated atomic facts and each candidate piece of evidence, including: For any candidate evidence, physical quantities are extracted from the text content corresponding to the candidate evidence to obtain the physical quantity constraint vector of the candidate evidence. For any constraint dimension in the physical quantity constraint vector, the degree of difference between the atomic fact and the candidate evidence under the constraint dimension is calculated based on the values ​​of the physical quantity constraint vector of the atomic fact and the physical quantity constraint vector of the candidate evidence under the constraint dimension. Based on the degree of difference between the atomic fact and the candidate evidence under different constraint dimensions in the physical quantity constraint vector, the degree of difference between the physical quantity of the atomic fact and the candidate evidence is obtained.

6. The method as described in claim 4, characterized in that, Based on the meteorological physics causal graph, the natural language description text and meteorological state vector corresponding to the node are corrected to obtain the corrected node, including: The atomic facts in the node whose verification result is not passed are taken as the target atomic facts; Based on the semantic relationship probability between the target atomic fact and each candidate piece of evidence, contradictory evidence is screened from multiple candidate pieces of evidence for the target atomic fact. Based on the contradictory evidence, the corrected natural language text of the node is generated, and the meteorological state vector of the node is updated according to the corrected natural language text to obtain the corrected node.

7. The method as described in claim 6, characterized in that, Based on the meteorological-physical causal graph, the natural language description text and meteorological state vector corresponding to the node are corrected to obtain the corrected node, which also includes: Based on the meteorological state vectors of the corrected node and the nodes downstream of the corrected node in the meteorological-physical causal graph, the semantic-physical consistency residual is calculated; If the semantic-physical consistency residual is greater than a preset residual threshold, then the natural language description text and meteorological state vector corresponding to the downstream node are corrected according to the meteorological-physical causal graph to obtain the corrected node.

8. A data synthesis device for training a meteorological large language model, characterized in that, The device includes: The acquisition unit is used to acquire meteorological questions and the corresponding standard answers to the meteorological questions; The deconstruction unit is used to reverse-construct the standard answer into a meteorological physics causal graph; the meteorological physics causal graph contains multiple nodes; each node corresponds to a natural language description text and a meteorological state vector; The verification unit is used to verify the natural language description text corresponding to each node according to the configured meteorological law evidence base, and obtain the verification result. The correction unit is used to correct the natural language description text and meteorological state vector corresponding to the node according to the meteorological physical cause-effect graph if the verification result of any node is not passed, so as to obtain the corrected node. The generation unit is used to generate the thought chain text based on the verification results of each node that passed the verification and the corresponding natural language description text of the corrected node. The determining unit is used to determine the thought chain text and the meteorological question as a set of thought chain data for training the meteorological agent.

9. An electronic device, characterized in that, The electronic device includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements the method of any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method described in any one of claims 1-7.

Images