Deduction method and system based on affair atlas and large model

Abstract

The invention belongs to the field of artificial intelligence, and provides a deduction method and system based on a affair atlas and a large model, which can expand the range of a data set by constructing event libraries containing real events and virtual events and establishing the affair atlas based on the event libraries, so that the data set is not limited by actual data, and the deduction efficiency is improved. Therefore, not only is the diversity of data enriched, but also the possibility of innovation is provided for the recommendation system, so that the recommendation result is richer and more diversified; an instruction data set for large model training is constructed by using a affair graph, and a complex relationship and a dynamic process between events can be better described, so that more accurate and in-depth knowledge support can be provided during construction of the instruction data set; the controllability and predictability of the output of the large model are improved through instruction tuning, namely, the output of an instruction constraint large model in an instruction data set; the large model has strong deduction capability and breaks through the scale limitation of the affair atlas database, and the deduction result may have certain innovativeness.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and in particular to a deduction method and system based on a causal graph and a large model. Background Art

[0002] The event reasoning task refers to predicting reasonable events that may occur in the future based on a known sequence of events. This task has important reference value in practical applications such as event risk prediction and event evolution analysis. In the prior art, the patent No. 202410049093.0 discloses a public opinion deduction method based on event graphs and related devices. The patent adopts a reasoning technology based on the event graph, and its effectiveness depends on the breadth of the event graph, the types of events covered, and the richness of the relationship between events. Therefore, the reasoning method proposed in the application cannot recommend event knowledge beyond the scope of the event graph library. In addition, the method continuously searches and matches in the knowledge base to determine the subsequent events that may occur, which leads to an extension of processing time and a decrease in matching accuracy. In summary, the deduction method in the current technology relies on the existing knowledge base for reasonable reasoning, which makes accurate search and timely matching for deduction difficult. Summary of the invention

[0003] In view of the problems that the deduction method in the prior art relies on the existing knowledge base and cannot perform reasonable reasoning without the existing knowledge base, and the search and matching process is time-consuming and has a low accuracy rate, the present invention proposes a deduction method and system based on a causal graph and a large model, which can quickly and accurately perform effective deduction based on events, and can break through the scale limitation of the knowledge base, and the deduction results have certain innovation.

[0004] A deduction method based on a causal graph and a large model, including: S1: Event data collection: collect open source data sets from the Internet, historical data from virtual simulations, and event texts from our own databases; S2: Constructing a causal graph: extracting the event text collected in S1 and classifying it into different fields, obtaining the event elements of virtual events and real events respectively, and storing them in the corresponding event library, and then further constructing a causal graph of the corresponding event based on the logical relationship between the event elements; S3: Construct instruction data set: construct an instruction data set for deduction tasks based on S2 event graph; each deduction instruction in the instruction data set is constructed by a logical event pair in the event graph, and the structure of the deduction instruction is: “Description of instruction: Based on the event background, deduce the subsequent behavior of the stakeholders of the cause event, and add the deduction results as the result event to the output item; Stakeholders: subjects and objects related to the event; Event background: relevant background information of the cause event; Input: Causing event; Output item: result event;" S4: Instruction tuning of the big model and evaluation: Use the S3 instruction data set to fine-tune the big model, obtain a big model that can perform deduction tasks and evaluate it, and execute subsequent reasoning tasks after meeting the evaluation standards; the model fine-tuning is based on the instruction data set to guide the big model to perform event reasoning; the big model automatically splices the input instruction description, stakeholders, event background and input items, and uses the spliced ​​content as the final input of the big model, and obtains the event deduction result of the corresponding input as the output of the big model, that is, the result event.

[0005] Preferably, the historical data of the virtual simulation in S1 is a simulation text history record generated in the virtual simulation; the simulation text history record records the response measures of the red and blue parties under different simulation backgrounds.

[0006] Preferably, the event elements described in step S2 include: a brief description of the event, a subject, an organization to which the subject belongs, an object, an organization to which the object belongs, the time of the event, the time at which the event ends, the location of the event, event verbs, the cause of the event, the result of the event, people related to the event, organizations related to the event, facilities related to the event, and event keywords.

[0007] Preferably, in step S2, the event elements are classified into a knowledge system of an event library according to categories, and the knowledge system of the event library includes a brief description of the event, a subject, an object, a time, a place, and an event classification.

[0008] Preferably, the method for extracting event text to obtain event elements in S2 is: A1: Construct prompt words that guide the big model to extract event elements from the event text; input the prompt words and event text into the big model to obtain the event elements in the corresponding input event text; A2: Manually verify the extracted event elements; construct an event element extraction training set using the prompt words constructed in A1, event text, and manually verified event elements; A3: Use the event element extraction training set to fine-tune the instructions of the large model to obtain the event extraction large model; A4: Use the event extraction model to extract event elements from the collected event text data.

[0009] Preferably, the specific method for classifying fields is: B1: Perform word segmentation and part-of-speech tagging on the event text to obtain the processed event text; B2: Use the pre-trained natural language processing model to perform feature extraction and word vector conversion on the processed event text to obtain the text vector of key text features; B3: The text vector of the key text feature is sent to the classifier. The classifier calculates the distance between the text vector and the classification template to determine the domain of the event.

[0010] Preferably, the method for further constructing a matter graph based on the logical relationship of the event elements in S2 is: C1, build an event library: the event library includes: a virtual event library built based on historical data of virtual deduction and a real event library built based on open source data sets collected from the Internet and a self-owned database; C2: Constructing logical relationships between events: Adding logical relationships between events based on the event library; the logical relationships include causal relationships and sequential relationships; C3: Construct a graph of events: Construct logical pairs based on the logical relationship and event library described in C2, and construct a graph of events in the form of logical event pairs.

[0011] Preferably, the process of constructing the event graph of the virtual event library is: a1: Extract the time, location, subject, object, position and brief description of the event from the historical data of virtual deduction; and classify the event into fields; the field classification includes politics, economy, military, diplomacy and society; the subject and object include our side, friendly side and enemy side; a2: Construct a virtual event library based on the extracted event time, event location, subject, object, position, event field classification and event brief description; a3: Use natural language processing methods to determine the logical relationship between events in the virtual event library constructed in a2; a4: Construct logical event pairs based on the logical relationship of a3 events and the virtual event library constructed in a2 to form a fact map.

[0012] Preferably, the process of constructing the event graph of the real event library is: b1: Extract the basic information of meta-events from event texts and classify the events by domain; the event texts are sourced from open source data sets on the Internet and proprietary databases; the basic information of meta-events includes brief descriptions of events, subjects, objects, time of events, and locations of events; the logical relationships between events are extracted using natural language processing methods, and the logical relationships between events include causality, sequence, condition, transition, parallelism, and co-occurrence; b2: Filter the event text according to the event domain classification and logical relationship obtained in b1; the information screening is to filter the events that are not in the domain classification, non-causal and non-sequential relationship; b3: Construct a real event database through the basic information of the screened meta-events and the field classification of events; b4: Construct logical event pairs based on the logical relationship between the real event library constructed by b3 and the events extracted by b1 to form a matter-of-fact graph.

[0013] Preferably, the construction method of each deduction instruction in the instruction data set in step S3 is: D1: Extract logical event pairs (A, B) from the logical graph; D2: Put the cause event A into the "Input:" item, and combine it into "Input: A"; put the result event B into the "Output:" item, and combine it into: "Output: B"; put the subjects and objects related to events A and B into the "[[Stakeholders]]" section; D3: For real events: extract a set of search terms from the cause event A to construct a search formula, retrieve relevant information from the constructed real event library as background information, and fill it into the "[[Event background]]" part; for virtual events, fill the background information of the virtual deduction in the constructed virtual event library into the "[[Event background]]" part; D4: Fill the event type of result event B into the "[[Field]]" part of the instruction description; D5: Fill the event subject of result event B into the "[[Related parties]]" part of the instruction description; D6: The structure of the deduced instructions in the final instruction data set is: Instruction description: Based on the event background, deduce the subsequent behavior of the stakeholders of the cause event, and add the deduction results as the result event to the output item; Stakeholders: subjects and objects related to the event; Event background: relevant background information of the cause event; Input: Causing event; Output item: result event.

[0014] A deduction system based on a causal graph and a large model, including: The instruction input module includes: an instruction description input unit for inputting the content of the "instruction description:" part; an input unit for inputting the content of the "input:" part; a stakeholder input unit for inputting the "[[stakeholders]]" part; and an event background input unit for inputting the "[[event background]]" part; Instruction merging module: used to splice the output contents of the instruction description input unit, input unit, stakeholder input unit and event background input unit to obtain a complete instruction; The large model module includes: a large model architecture selection unit for selecting large models of different architecture types; a result output unit for inputting the complete instruction synthesized by the instruction synthesis module into the large model selected by the large model architecture selection unit, and obtaining the deduction result corresponding to the input instruction; all large models in the large model architecture selection unit have completed instruction tuning, meet the evaluation criteria, and have the ability to perform deduction tasks; The display module is used to put the deduction results of the result output unit into the output item of the instruction data set, namely the "output: deduction results" part for display.

[0015] Beneficial effects: The present invention proposes a deduction method and system based on event graph and big model. By using event graph to construct instruction data set for big model training, the big model is fine-tuned so that the big model can perform deduction tasks. The pre-trained big model stores a lot of knowledge in the real world. After instruction tuning training, the big model has a strong deduction ability. With the help of the massive knowledge stored in itself, it can make better recommendations, breaking through the scale limit of the event graph library, and the deduction results may have certain innovations.

[0016] By building a large event element extraction model, we can quickly and accurately extract basic information about the event from the event text; when classifying event text, word segmentation helps to divide continuous text into meaningful units, making it easier to extract important information such as keywords and entities; part-of-speech tagging can provide basic information for syntactic analysis and help determine the grammatical relationship between words in a sentence, so as to better understand the structure and semantics of the sentence, thereby improving the accuracy of event text field classification.

[0017] By constructing an event library containing real events and virtual events, and establishing a causal map based on these event libraries, the scope of the data set can be expanded so that it is not restricted by actual data. This not only enriches the diversity of the data, but also provides innovative possibilities for the reasoning model, making the deduction results richer and more diverse. By constructing a causal map, the causal and sequential relationships between events can be emphasized, and the complex relationships and dynamic processes between events can be better described, thereby providing more accurate and in-depth knowledge support in constructing the instruction data set.

[0018] The large model has stronger representation and generalization capabilities, can demonstrate higher accuracy and robustness in the reasoning process, and can give deduction results independent of the existing knowledge base, and the deduction results have a certain degree of rationality and accuracy; at the same time, the large model can capture subtle features in the data, so that when dealing with complex and detailed reasoning tasks, it can better understand and process the instruction descriptions and inputs in the instruction data set, thereby obtaining more accurate and reasonable deduction results.

[0019] Through instruction tuning, the instructions in the instruction data set constrain the output of the large model to keep it consistent with the required domain knowledge, and improve the controllability and predictability of the output of the large model; instruction tuning can help the large model quickly adapt to reasoning tasks without a lot of retraining or architectural changes. The large model is usually trained in value alignment and other training. After instruction tuning technology, the large model has a strong ability to follow deduction instructions, which greatly reduces the probability of refusing to answer deduction questions. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a flowchart of a deduction method based on the causal map and the large model.

[0021] Figure 2 It is a structural diagram of a deduction system based on the graph of facts and the large model. DETAILED DESCRIPTION

[0022] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0023] like Figure 1 As shown, the specific process of a deduction method based on the event map and the large model is as follows: S1: Event data collection: collect open source data sets from the Internet, historical data from virtual simulations, and event texts from our own databases; S2: Constructing a causal graph: extracting the event text collected in S1 and classifying it into different fields, obtaining the event elements of virtual events and real events respectively, and storing them in the corresponding event library, and then further constructing a causal graph of the corresponding event based on the logical relationship between the event elements; S3: Construct instruction data set: construct an instruction data set for deduction tasks based on S2 event graph; each deduction instruction in the instruction data set is constructed by a logical event pair in the event graph, and the structure of the deduction instruction is: “Description of instruction: Based on the event background, deduce the subsequent behavior of the stakeholders of the cause event, and add the deduction results as the result event to the output item; Stakeholders: subjects and objects related to the event; Event background: relevant background information of the cause event; Input: Causing event; Output item: result event;" S4: Instruction tuning of the big model and evaluation: Use the S3 instruction data set to fine-tune the big model, obtain a big model that can perform deduction tasks and evaluate it, and execute subsequent reasoning tasks after meeting the evaluation standards; the model fine-tuning is based on the instruction data set to guide the big model to perform event reasoning; the big model automatically splices the input instruction description, stakeholders, event background and input items, and uses the spliced ​​content as the final input of the big model, and obtains the event deduction result of the corresponding input as the output of the big model, that is, the result event.

[0024] The specific implementation process is: 1. A deduction method based on a causal graph and a large model includes the following steps: First, based on open source data collected from the Internet or historical data based on virtual deduction, basic event information and the logical relationship between events are extracted to construct a cause-and-effect map.

[0025] Based on the event graph, construct an instruction dataset for event deduction tasks; Use the instruction dataset to fine-tune the pre-trained large model; Apply tuned large models to perform event simulations.

[0026] 2. The construction method of the event map: Based on the historical information of virtual deduction, the basic information of events and the logical relationship between events are extracted to build a virtual event library; Based on open source data collected from the Internet and its own knowledge base, it extracts events and the logical relationships between them to build a real event database; The event information extracted includes: the event field classification (including politics, economy, military, diplomacy, etc.), event subject and object (including country, organization, person, etc.), brief description information, and logical relationships such as cause and effect and sequence between events; The method to extract event information is: Step 1: Define prompt words and use the big model capabilities to directly extract event elements from the document.

[0027] The prompt word instructions are as follows: Now you are an intelligence analyst, and your task is to extract events and event elements (events) from the input document. The extracted event elements include: brief description of the event, subject, agency to which the subject belongs, object, agency to which the object belongs, time of event occurrence, time of event end, location of event occurrence, event verb, cause of event occurrence, result of event occurrence, people related to the event, agency related to the event, facilities related to the event, and event keywords.

[0028] Step 2: Manually verify the extraction results, optimize the extraction results, and form an event element extraction training set.

[0029] Step 3: Fine-tune the instructions of the large model and train the event extraction large model.

[0030] Step 4: Use the event extraction model to extract event elements from the data.

[0031] (1) The process of constructing a causal map based on historical records: Historical game records are game text records generated by users during the virtual game process. These data record the response events of the red and blue parties under different game backgrounds.

[0032] Natural language processing technology is used to extract the subject and object of events (our side, friendly side, enemy side, etc.), country and position, and event classification (including politics, economy, military, diplomacy, society, etc.) from the simulation records.

[0033] The specific method of extracting causal and sequential relationships from the deduction records and judging logical relationships is as follows: 1) There are multiple parties in the simulation, including our side and the enemy side; 2) After we give a strategy A (event A), the enemy responds with strategy B (event B) within a certain period of time, then we can set up a causal relationship between strategy B and strategy A, and construct the enemy-us causal event pair (A, B); 3) After we give a strategy A (event A), we give a strategy C (event C) within a certain period of time. Then we can set a sequential relationship between strategy C and strategy A, and construct an I-I sequential event pair (A, C).

[0034] (2) The process of constructing a causal graph based on Internet news and own knowledge base Filter information from Internet news pages and own knowledge base based on pre-set topics; Natural language processing technology is used to extract meta-events and the relationship between events from texts filtered by topics such as web pages, and an Internet event database is constructed.

[0035] Extracting event information includes: event field classification (including politics, economy, military, diplomacy, society, etc.), event subject and object (including country, organization, person, etc.), event brief description, logical relationship between events such as cause and effect and sequence; extracting information such as country and position of both parties (subject and object) of the event; Event filtering: filter out events outside the domain classification; filter out events that are not causal and sequential; 3. Based on the event graph, construct an instruction dataset for deduction tasks: Specifically, the format of the instruction data is: { Instruction description: Based on the event background, predict the measures or statements that [[relevant parties]] will take in the [political, economic, military, diplomatic, etc.] fields in response to the event input by the user.

[0036] [[Stakeholders]]: xxxxxxxxx; [[Event background]]: xxxxxxxxx; Input: xxxxxx; Output: xxxxxxxx }; The instruction description is an instruction to guide the large model to conduct event reasoning on the relevant parties within the event field based on the event background and input content; the relevant parties are manually filled in based on the country information of the event subject; and the field is manually filled in with reference to the field classification information of the event.

[0037] The relevant parties are mainly generated based on the country information of the event subject.

[0038] The domain mainly refers to the domain classification information generation of the event.

[0039] For the above instruction data format, the method of constructing the deduction instruction is as follows: 1) Extract causal event pairs (A, B); 2) Put the cause event A into the "Input:" section; 3) Put the result event B into the "Output:" section; 4) Put the subjects and objects related to events A and B into the “[[Stakeholders]]” section; 5) For real events: extract a set of search terms from the cause event to construct a search formula, retrieve relevant information from the collection database as background information, and fill it into the "[[Event background]]" part; For virtual events, fill in the "[[Event Background]]" section with the background information of the virtual event; 6) Fill the event type of result event B into the "[field]" field; 7) Fill in the event body of result event B into the “[[Related parties]]” section.

[0040] 4. Fine-tune the large model based on the instruction dataset; The instruction dataset constructed based on the event graph is sent to a large model (such as LLAMA, ChatGLM, etc.) for fine-tuning training. The fine-tuning method can be LoRA, QLoRA, P-Tuning, etc.

[0041] 5. Provide deduction services based on large models.

[0042] After training, the large model is evaluated, which includes automatic evaluation and manual evaluation; the automatic evaluation is to create a test instruction data set with correct answers, and evaluate the large model through accuracy, precision and F1 score; the manual evaluation is to input the test instructions into the large model and manually judge whether the output result of the large model is correct.

[0043] The large model after evaluation can have a stronger event deduction capability.

[0044] The following command request can be constructed: { Instruction description: Based on the event background and the event input by the user, predict the measures or statements that Company A will take in the political, economic, military, diplomatic and other fields in response to the event.

[0045] [[Stakeholders]] Company A [[Background of the incident]] In order to enhance employees' work enthusiasm and improve work efficiency, Company A plans to purchase a large amount of office equipment.

[0046] Input: A large number of equipment purchased by Company A suddenly malfunctioned, which adversely affected the work efficiency of employees and the company's information security.

[0047] Output: } Submit the constructed instruction request to the big model for event deduction. The output of the big model is: { Output: 1. Notify maintenance personnel, supervisors and heads of relevant departments to ensure that they are aware of the situation and take appropriate measures in a timely manner.

[0048] 2. Regular inspection: Carry out regular and thorough inspections on communication equipment to prevent equipment failures and abnormalities.

[0049] 3. Operator training: Strengthen the training of equipment operators, improve their operating skills and safety awareness, ensure that operators can operate the equipment correctly, and avoid equipment failures caused by misoperation.

[0050] 4. Supplier verification: Company A will conduct strict reviews of all suppliers and sellers and select suppliers with higher product quality to avoid problems with large quantities of equipment again.

[0051] } like Figure 2 As shown, a deduction system based on a causal graph and a large model includes: The instruction input module includes: an instruction description input unit for inputting the content of the "instruction description:" part; an input unit for inputting the content of the "input:" part; a stakeholder input unit for inputting the "[[stakeholders]]" part; and an event background input unit for inputting the "[[event background]]" part; Instruction merging module: used to splice the output contents of the instruction description input unit, input unit, stakeholder input unit and event background input unit to obtain a complete instruction; The large model module includes: a large model architecture selection unit for selecting large models of different architecture types; a result output unit for inputting the complete instruction synthesized by the instruction synthesis module into the large model selected by the large model architecture selection unit, and obtaining the deduction result corresponding to the input instruction; all large models in the large model architecture selection unit have completed instruction tuning, meet the evaluation criteria, and have the ability to perform deduction tasks; The display module is used to put the deduction results of the result output unit into the output item of the instruction data set, namely the "output: deduction results" part for display.

[0052] It should be noted that the above-described specific implementations can enable those skilled in the art to more fully understand the invention, but do not limit the invention in any way. Therefore, although this specification has described the invention in detail with reference to the drawings and embodiments, those skilled in the art should understand that the invention can still be modified or replaced by equivalents. In short, all technical solutions and improvements that do not deviate from the spirit and scope of the invention should be included in the protection scope of the patent for the invention.

Claims

1. A deduction method based on a causal graph and a large model, characterized in that: S1: Event data collection: collect open source data sets from the Internet, historical data from virtual simulations, and event texts from our own databases; S2: Constructing a causal graph: extracting the event text collected in S1 and classifying it into different fields, obtaining the event elements of virtual events and real events respectively, and storing them in the corresponding event library, and then further constructing a causal graph of the corresponding event based on the logical relationship between the event elements; S3: Construct instruction data set: construct an instruction data set for deduction tasks based on S2 event graph; each deduction instruction in the instruction data set is constructed by a logical event pair in the event graph, and the structure of the deduction instruction is: "Instruction description: Based on the event background, deduce the subsequent behavior of the stakeholders of the cause event, and add the deduction results as the result event to the output item; Stakeholders: subjects and objects related to the event; Event background: relevant background information of the cause event; Input: Causing event; Output item: result event;" S4: Instruction tuning of the big model and evaluation: Use the S3 instruction data set to fine-tune the big model, obtain a big model that can perform deduction tasks and evaluate it, and execute subsequent reasoning tasks after meeting the evaluation standards; the model fine-tuning is based on the instruction data set to guide the big model to perform event reasoning; the big model automatically splices the input instruction description, stakeholders, event background and input items, and uses the spliced ​​content as the final input of the big model, and obtains the event deduction result of the corresponding input as the output of the big model, that is, the result event.

2. A deduction method based on a causal graph and a large model according to claim 1, characterized in that: The historical data of the virtual game described in S1 is the game text history record generated in the virtual game; the game text history record records the response measures of the red and blue parties under different game backgrounds.

3. A deduction method based on a causal graph and a large model according to claim 1, characterized in that: The event elements described in step S2 include: a brief description of the event, a subject, an organization to which the subject belongs, an object, an organization to which the object belongs, the time of the event, the time of the event ending, the location of the event, the event verb, the cause of the event, the result of the event, people related to the event, organizations related to the event, facilities related to the event, and event keywords.

4. A deduction method based on a causal graph and a large model according to claim 1, characterized in that: In step S2, the event elements are classified into a knowledge system of an event library according to various categories. The knowledge system of the event library includes a brief description of the event, a subject, an object, a time, a place, and an event classification.

5. A deduction method based on a causal graph and a large model according to claim 1, characterized in that: The method for extracting event text to obtain event elements in S2 is: A1: Construct prompt words that guide the big model to extract event elements from the event text; input the prompt words and event text into the big model to obtain the event elements in the corresponding input event text; A2: Manually verify the extracted event elements; construct an event element extraction training set using the prompt words constructed in A1, event text, and manually verified event elements; A3: Use the event element extraction training set to fine-tune the instructions of the large model to obtain the event extraction large model; A4: Use the event extraction model to extract event elements from the collected event text data.

6. A deduction method based on a causal graph and a large model according to claim 1, characterized in that: The specific method for field classification is as follows: B1: Perform word segmentation and part-of-speech tagging on the event text to obtain the processed event text; B2: Use the pre-trained natural language processing model to perform feature extraction and word vector conversion on the processed event text to obtain the text vector of key text features; B3: The text vector of the key text feature is sent to the classifier. The classifier calculates the distance between the text vector and the classification template to determine the domain of the event.

7. A deduction method based on a causal graph and a large model according to claim 1, characterized in that: The method for further constructing the event graph based on the logical relationship of the event elements in S2 is: C1, build an event library: the event library includes: a virtual event library built based on historical data of virtual deduction, and a real event library built based on open source data sets collected from the Internet and its own database; C2: Constructing logical relationships between events: Adding logical relationships between events based on the event library; the logical relationships include causal relationships and sequential relationships; C3: Construct a graph of events: Construct logical pairs based on the logical relationship and event library described in C2, and construct a graph of events in the form of logical event pairs.

8. A deduction method based on a causal graph and a large model according to claim 7, characterized in that: The process of constructing the event graph of the virtual event library is as follows: a1: Extract the time, location, subject, object, position and brief description of the event from the historical data of virtual deduction; and classify the event into fields; the field classification includes politics, economy, military, diplomacy and society; the subject and object include our side, friendly side and enemy side; a2: Construct a virtual event library based on the extracted event time, event location, subject, object, position, event field classification and event brief description; a3: Use natural language processing methods to determine the logical relationship between events in the virtual event library constructed in a2; a4: Construct logical event pairs based on the logical relationship of a3 events and the virtual event library constructed in a2 to form a fact map.

9. A deduction method based on a causal graph and a large model according to claim 7, characterized in that: The process of constructing the event graph of the real event library is as follows: b1: Extract the basic information of meta-events from event texts and classify the events by domain; the event texts are sourced from open source data sets on the Internet and proprietary databases; the basic information of meta-events includes brief descriptions of events, subjects, objects, time of events, and locations of events; the logical relationships between events are extracted using natural language processing methods, and the logical relationships between events include causality, sequence, condition, transition, parallelism, and co-occurrence; b2: Filter the event text according to the event domain classification and logical relationship obtained in b1; the information screening is to filter the events that are not in the domain classification, non-causal and non-sequential relationship; b3: Construct a real event database through the basic information of the screened meta-events and the field classification of events; b4: Construct logical event pairs based on the logical relationship between the real event library constructed by b3 and the events extracted by b1 to form a matter-of-fact graph.

10. A deduction method based on a causal graph and a large model according to claim 1, characterized in that: The construction method of each deduction instruction in the instruction data set in step S3 is: D1: Extract logical event pairs (A, B) from the logical graph; D2: Place the cause event A in the "Input:" section to form "Input: A"; place the result event B in the "Output:" section to form "Output: B"; place the subjects and objects related to events A and B in the "[[Stakeholders]]" section; D3: For real events: extract a set of search terms from the cause event A to construct a search formula, retrieve relevant information from the constructed real event library as background information, and fill it into the "[[Event background]]" part; for virtual events, fill the background information of the virtual deduction in the constructed virtual event library into the "[[Event background]]" part; D4: Fill the event type of result event B into the "[[Field]]" part of the instruction description; D5: Fill the event subject of result event B into the "[[Related parties]]" part of the instruction description; D6: The structure of the deduced instructions in the final instruction data set is: Instruction description: Based on the event background, deduce the subsequent behavior of the stakeholders of the cause event, and add the deduction results as the result event to the output item; Stakeholders: subjects and objects related to the event; Event background: relevant background information of the cause event; Input: Causing event; Output item: result event.

11. A deduction system based on a factual graph and a large model formed based on the method described in any one of claims 1 to 10, characterized in that: include: The instruction input module includes: an instruction description input unit for inputting the content of the "instruction description:" part; an input unit for inputting the content of the "input:" part; a stakeholder input unit for inputting the "[[stakeholders]]" part; and an event background input unit for inputting the "[[event background]]" part; Instruction merging module: used to splice the output contents of the instruction description input unit, input unit, stakeholder input unit and event background input unit to obtain a complete instruction; The large model module includes: a large model architecture selection unit for selecting large models of different architecture types; a result output unit for inputting the complete instruction synthesized by the instruction synthesis module into the large model selected by the large model architecture selection unit, and obtaining the deduction result corresponding to the input instruction; all large models in the large model architecture selection unit have completed instruction tuning, meet the evaluation criteria, and have the ability to perform deduction tasks; The display module is used to put the deduction results of the result output unit into the output item of the instruction data set, namely the "output: deduction results" part for display.

Images