An event evolution reasoning and prediction method based on cognitive-driven intelligent generation

Abstract

This application discloses an event evolution reasoning and prediction method based on cognitive-driven intelligent generation. First, it utilizes a domain-wide language model to extract expert knowledge points, analysis processes, analytical perspectives, and logical reasoning from event research reports, constructing an interactive thought chain. It then extracts causal reasoning and constructs a causal reasoning network structure. Cognitive logical knowledge is integrated with relevant facts of the event to be predicted and embedded into a prompt template to construct a prompt model. This allows the domain-wide language model to generate event evolution reasoning and prediction results based on expert analysis dimensions and processes. Finally, conflict detection is performed on event elements through closed-loop evidence chain analysis, and credible conclusions are selected to obtain credible prediction conclusions. By constructing prompt templates to guide the large model through cognitive logical learning and combining multi-level credible verification and screening, the technical problem of excessive prediction errors caused by illusions generated by the large model is solved, improving the accuracy and credibility of event evolution prediction.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and more specifically, to an event evolution reasoning and prediction method based on cognitively driven intelligent generation. Background Technology

[0002] Event evolution prediction technology is a key area for decision-making in complex systems. It aims to assist in handling highly dynamic and uncertain multi-agent interactive problems by modeling, extrapolating, and predicting event development trends. This technology typically relies on methods such as knowledge representation, logical reasoning, and data analysis to construct event evolution models to support the decision-making process, and has significant application value in scenarios such as emergency management and social situation analysis.

[0003] However, traditional methods for predicting event evolution have significant limitations in practice. These methods often require pre-set conclusions or rely on fixed rules, and their predictions are constrained by the framework of prior human cognition, making them ineffective in handling unforeseen or sudden events. Because they cannot fully characterize the nonlinear relationships and dynamic evolution mechanisms in complex systems, traditional methods lack accuracy and adaptability when dealing with open and uncertain environments, and are unable to provide forward-looking guidance.

[0004] In recent years, with the development of computing technology, open heuristic-based prediction methods have emerged. By deeply analyzing the internal logic and relationship networks of events, these methods enhance the insight into evolutionary mechanisms and, to some extent, surpass traditional statistical methods. However, these methods still have reliability issues when generating prediction conclusions, and are prone to producing inferences that do not conform to the facts, affecting the effectiveness of decision-making. Therefore, improvements are urgently needed to enhance the robustness and practicality of predictions. Summary of the Invention

[0005] The purpose of this application is to overcome the shortcomings of existing technologies and provide an event evolution reasoning and prediction method based on cognitively driven intelligent generation. This purpose is achieved through the following technical solution: Firstly, this application proposes an event evolution reasoning prediction method based on cognitively driven intelligent generation, including: By using a domain-wide language model, expert knowledge points, analysis processes, analytical perspectives, and logical reasoning are extracted from event research reports to construct an interactive thinking chain. Based on the event research reports, a domain event knowledge base is built to extract causal reasoning and construct a causal reasoning network structure. Based on the interactive thinking chain and causal reasoning network structure, cognitive logic knowledge and problem-related facts of the event to be predicted are integrated and embedded into the prompt template to construct a prompt model. This enables the domain-wide language model to generate event evolution reasoning prediction results based on expert analysis dimensions and analysis processes. Based on the event evolution reasoning prediction results and the related information vector retrieval results, conflict detection is performed on event elements through evidence chain closed-loop analysis, and credible conclusions are screened through a false detection model to obtain credible prediction conclusions.

[0006] In one possible implementation, the steps of extracting expert knowledge points, analysis processes, analytical perspectives, and logical reasoning from event research reports using a domain-specific large language model to construct an interactive thought chain include: Sentences from the event investigation report text Assign chapter element tags The elements of a text include an introduction, a central argument, sub-arguments, factual evidence, theoretical evidence, and a conclusion. Based on the identified text elements, structured extraction is performed to build an interactive thought chain, outputting problem analysis perspectives, analysis processes, and knowledge points.

[0007] In one possible implementation, the steps of constructing a domain event knowledge base based on event research reports, extracting causal principles, and constructing a causal principle network structure include: A method combining knowledge enhancement and prompting learning is adopted. A domain event knowledge base is built based on event research reports. Relevant domain event knowledge is injected into prompt templates and input into a general large language model to generate structured causal reasoning extraction results. Based on the results of causal reasoning extraction, a causal reasoning network structure is constructed. ,in For a set of event nodes, It is the set of edges representing causal relationships between events.

[0008] In one possible implementation, the steps of constructing a prompt model by integrating cognitive logical knowledge with problem-related facts of the event to be predicted into a prompt template based on an interactive thought chain and causal reasoning network structure include: Based on cognitive knowledge, the task of predicting the development trend of events is analyzed and understood, the task type is identified, and key elements of the event background in the task text are extracted. The reasoning and prediction steps for constructing corresponding tasks based on interactive thinking chains ; Based on the task description instructions, inference and prediction steps, input task question sentences, and output format requirements, a prompt template is constructed. ,in For the general construction of prompt templates, For task description instructions, For task reasoning prediction steps , To input the task question sentence, For output format requirements.

[0009] In one possible implementation, the following is included before constructing the prompt template: Based on cognitive knowledge, the task of predicting the development trend of events is analyzed and understood, the task type is identified, and key elements of the event background in the task text are extracted.

[0010] In one possible implementation, the process of constructing the prompt template further includes: The generative language model is taught using manually decomposed examples so that it can learn task-related domain contextual knowledge from the examples.

[0011] In one possible implementation, the step of performing conflict detection on event elements through closed-loop evidence chain analysis based on event evolution reasoning prediction results and associated information vector retrieval results includes: Vector similarity retrieval technology is used to retrieve recent key information related to the prediction problem from background information as the associated information vector retrieval results; Event elements are extracted from related information, and conflict detection is performed on the event evolution reasoning prediction results based on content detection rules from three aspects: spatiotemporal conflict, target attribute conflict, and content logic conflict.

[0012] In one possible implementation, conflict detection calculates a comprehensive conflict score using a weighted summation method. ,in Due to a conflict in time and space, Due to a conflict in target attributes, Due to logical conflicts in the content, All are weighting coefficients.

[0013] In one possible implementation, the step of filtering credible conclusions using a false detection model to obtain credible prediction conclusions includes: A false alarm template is constructed by linking information and inference predictions. The inference predictions are then filtered using a false alarm detection model, retaining those that pass the false alarm detection, and the inference predictions are ranked by their credibility. ,in For the false detection model to correlate information The detection rules As weight.

[0014] In one possible implementation, the method further includes: The system sorts the multiple predictions that pass the screening by credibility and outputs a list of the sorted credible predictions.

[0015] The main solution and its various further alternatives described above can be freely combined to form multiple solutions, all of which are solutions that can be adopted and are claimed in this application; furthermore, the (non-conflicting alternatives) can also be freely combined with each other and with other alternatives. Those skilled in the art, after understanding the solution of this application, will realize from the prior art and common general knowledge that there are many combinations, all of which are technical solutions to be protected in this application, and will not be exhaustively listed here.

[0016] This application discloses an event evolution reasoning and prediction method based on cognitive-driven intelligent generation. First, it utilizes a domain-wide language model to extract expert knowledge points, analysis processes, analytical perspectives, and logical reasoning from event research reports, constructing an interactive thought chain. It then extracts causal reasoning and constructs a causal reasoning network structure. Cognitive logical knowledge is integrated with relevant facts of the event to be predicted and embedded into a prompt template to construct a prompt model. This allows the domain-wide language model to generate event evolution reasoning and prediction results based on expert analysis dimensions and processes. Finally, conflict detection is performed on event elements through closed-loop evidence chain analysis, and credible conclusions are selected to obtain credible prediction conclusions. By constructing prompt templates to guide the large model through cognitive logical learning and combining multi-level credible verification and screening, the technical problem of excessive prediction errors caused by illusions generated by the large model is solved, improving the accuracy and credibility of event evolution prediction. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments 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.

[0018] Figure 1 The diagram shows a flowchart of an event evolution reasoning prediction method based on cognitive-driven intelligent generation proposed in an embodiment of this application.

[0019] Figure 2 Another flowchart of the event evolution reasoning prediction method proposed in this application is shown. Detailed Implementation

[0020] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0021] Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] To address the problem of excessive errors in event evolution reasoning prediction techniques due to the illusion generated by large models, this application proposes an event evolution reasoning prediction method based on cognitively driven intelligent generation. By employing a cognitive logic knowledge learning method, it extracts expert knowledge points, analysis processes, and analytical perspectives, and constructs an interactive logical thinking chain by combining causal reasoning extraction. Under cognitive drive and combined with large model retrieval enhancement technology, it generates logically consistent prediction conclusions. Furthermore, it utilizes text false detection technology to verify the authenticity of multiple generated conclusions, thereby improving the credibility of the large model's prediction conclusions.

[0023] Please refer to Figure 1 , Figure 1 This paper illustrates a flowchart of an event evolution reasoning prediction method based on cognitively driven intelligent generation, as proposed in an embodiment of this application, including: Step S1: Use the domain big language model to extract expert knowledge points, analysis process, analysis perspective and logical reasoning from the event research report, construct an interactive thinking chain, and build a domain event knowledge base based on the event research report, extract causal reasoning and construct a causal reasoning network structure.

[0024] In the expert thinking learning process, the system uses event research reports or papers in relevant professional fields as input. Utilizing a domain-wide language model, it transforms the expert thinking extraction task into a sequence labeling task, automatically assigning a predefined discourse element label to each sentence in the text. By analyzing the argumentative logic represented by these structured labels, the system can automatically extract the expert's analytical process, perspective, and key knowledge points from the text, thereby constructing an interactive thought chain reflecting the expert's reasoning path. In the causal reasoning extraction process, the system employs a combination of knowledge enhancement and prompting learning: First, a structured domain event knowledge base is constructed based on relevant event research reports within the domain; then, the domain knowledge from this knowledge base is injected into a designed prompt template and input into the general language model to guide the model in extracting structured cause-effect pairs from the text. Finally, based on these extraction results, a formalized causal reasoning network structure is constructed.

[0025] Step S2: Based on the interactive thinking chain and causal reasoning network structure, integrate cognitive logic knowledge with the problem-related facts of the event to be predicted and embed them into the prompt template to construct a prompt model. This enables the domain-wide language model to generate event evolution reasoning prediction results based on expert analysis dimensions and analysis processes.

[0026] First, the acquired cognitive logic knowledge—namely, the interactive thought chain extracted from expert thinking learning and the causal logic network structure constructed from causal logic extraction—is deeply integrated with the specific problem-related facts of the event to be predicted. These three types of knowledge are collectively embedded in a carefully designed prompt template. By constructing this prompt model, the domain-wide language model is guided to perform step-by-step reasoning based on the expert analysis framework and causal logic chain embedded in the template, ultimately generating standardized and structured preliminary event evolution reasoning and prediction results.

[0027] Step S3: Based on the event evolution reasoning prediction results and the associated information vector retrieval results, conflict detection is performed on the event elements through evidence chain closed-loop analysis, and credible conclusions are screened through a false detection model to obtain credible prediction conclusions.

[0028] Based on the results of the associated information vector retrieval, this retrieval utilizes vector similarity technology to obtain recent key information related to the prediction question from the background information database. Next, event element conflict detection is performed, extracting structured event elements such as time, location, people, cause, purpose, and result from the retrieved associated information. These elements are then compared with the generated event evolution reasoning prediction results. According to preset rules, a closed-loop analysis of the evidence chain is conducted from three aspects: spatiotemporal conflict, target attribute conflict, and content logic conflict, and a comprehensive conflict score is calculated. For false detection, a specific prompt template is constructed from the associated information and the prediction conclusion to be verified. This template is input into the false detection model for authenticity judgment, filtering out conclusions judged as false. For conclusions that pass the detection, their credibility score is calculated. Finally, all prediction conclusions that pass the filtering are ranked according to their credibility scores, and a ranked list of credible prediction conclusions is output.

[0029] The steps for extracting expert knowledge points, analysis processes, analytical perspectives, and logical reasoning from event research reports using a domain-specific large language model to construct an interactive thought chain include: Sentences from the event investigation report text Assign chapter element tags The elements of a text include an introduction, a central argument, sub-arguments, factual evidence, theoretical evidence, and a conclusion. Based on the identified text elements, structured extraction is performed to build an interactive thought chain, outputting problem analysis perspectives, analysis processes, and knowledge points.

[0030] In the expert thinking learning process, this method first treats the text of an event research report or professional paper as a sequence of sentences. Next, through the text element recognition task, a predefined text element label is automatically assigned to each sentence, thus forming a corresponding label sequence. These text element tags specifically include: Introduction (identifying sentences that introduce the background of events or lay the groundwork for the main idea); Central Argument (identifying sentences that express the author's core claim); Sub-arguments (identifying finer-grained claims that support the central argument); Factual Evidence (identifying sentences that state actual events or data to support the argument); Theoretical Evidence (identifying sentences that provide theoretical explanations or basis); and Conclusion (identifying sentences that summarize the entire text and echo the main idea). Sentences that do not fall into any of these categories are marked as "Other."

[0031] After assigning text element tags to all sentences, the system extracts structured information based on these identification results. It constructs an interactive thought chain reflecting the expert's reasoning path by analyzing the supporting relationships between the central argument and sub-arguments, the evidence chain relationships between sub-arguments and factual / theoretical arguments, and the overall argumentative flow from introduction to conclusion. Ultimately, this process outputs structured cognitive knowledge, mainly including: the problem analysis perspective extracted from the argumentative framework, the problem analysis process summarized from the argumentative order, and key knowledge points extracted from various arguments. The text element definitions and classifications are shown in Table 1. Table 1

[0032] The steps involved in building a domain event knowledge base based on event research reports, extracting causal principles, and constructing a causal principle network structure include: A method combining knowledge enhancement and prompting learning is adopted. A domain event knowledge base is built based on event research reports. Relevant domain event knowledge is injected into prompt templates and input into a general large language model to generate structured causal reasoning extraction results. Based on the results of causal reasoning extraction, a causal reasoning network structure is constructed. ,in For a set of event nodes, It is the set of edges representing causal relationships between events.

[0033] First, a structured domain event knowledge base is constructed based on research reports on relevant thematic events within the domain. Next, the relevant domain event knowledge from the knowledge base is combined with causal extraction instructions and injected into a carefully designed prompt template. This template aims to guide a large language model to recognize causal relationships in text. Then, the prompt template containing domain knowledge and the text to be processed are input into a general large language model. Based on the injected knowledge and prompt instructions, the model extracts structured causal pairs from the text. Finally, based on these extracted structured causal pairs, a causal network graph is constructed. Among them, the set This represents the individual event nodes extracted from the text, while the set... It is formed by directed edges The set consists of elements that exist in the event nodes. and The causal relationship between them.

[0034] Based on an interactive thinking chain and causal network structure, the steps for constructing a prompt model by integrating cognitive logical knowledge with problem-related facts of the event to be predicted into a prompt template include: Based on cognitive knowledge, the task of predicting the development trend of events is analyzed and understood, the task type is identified, and key elements of the event background in the task text are extracted. The reasoning and prediction steps for constructing corresponding tasks based on interactive thinking chains ; Based on the task description instructions, inference and prediction steps, input task question sentences, and output format requirements, a prompt template is constructed. ,in For the general construction of prompt templates, For task description instructions, For task reasoning prediction steps , To input the task question sentence, For output format requirements.

[0035] By combining logical cognition knowledge to understand reasoning and prediction tasks, three types of knowledge—expert thinking, causal reasoning, and problem-related facts—are embedded into the prompt template to construct a prompt model. This enables the large language model to generate standardized event evolution reasoning and prediction results based on the expert analysis dimensions and analysis process.

[0036] Given that a general large language model is used, in order to further clarify the output position and format of the model, the constructed prompt template adopts a "cloze test" structure, that is, clear blanks or marks are reserved in the template to guide the model to generate standardized content that meets the requirements in the specified position.

[0037] The structure of the prompt template is as follows: 1) Task Instruction: This is a [ ] task. For the input [ ] instruction, perform the [ ] operation and output the [ ] content, such as [ ] (detailed example, optional).

[0038] 2) Input sentence: So what is the output result for the problem 【 】?

[0039] 3) Output requirements: The result should be returned in the format of [ ] (specify an easy-to-parse format such as json / xml).

[0040] The task instructions are fixed prompt templates generated based on three parts of knowledge: expert thinking, causal reasoning, and relevant facts. Different templates are pre-generated for different event evolution reasoning and prediction tasks. The reasoning and prediction steps for the corresponding tasks are constructed based on the interactive reasoning chain (CoT). , Indicates solving the task The required first A specific reasoning step.

[0041] In summary, the standard structural expression for a task prompt template is: Where T is the general construct for the prompt template, I is the task description instruction, and C is the task reasoning and prediction step. D represents the input task question, and F represents the output format requirements.

[0042] The general-purpose large language model analyzes and internally infers from task prompt templates and manually decomposed examples, and combines relevant factual data from the domain event knowledge base to generate intelligent reasoning and guessing results for the input question: ,in This is the task prompt template you entered. This is the parameterized generation function for large models.

[0043] The construction prompt template also includes: Based on cognitive knowledge, the task of predicting the development trend of events is analyzed and understood, the task type is identified, and key elements of the event background in the task text are extracted.

[0044] The system identifies and categorizes task types. Based on the task's objectives and characteristics, it classifies tasks into a predefined task type system. For example, it identifies whether the task requires deducing multiple possible paths for the future development of an event, or making a comprehensive judgment on the overall situation or final outcome of the event. Secondly, the system extracts key elements of the event background from the task text. It automatically identifies and extracts core elements such as time, location, event subject, event object, and related actions from the text describing the prediction problem.

[0045] When constructing the prompt template, it also includes: The generative language model is taught using manually decomposed examples so that it can learn task-related domain contextual knowledge from the examples.

[0046] Based on three types of knowledge—expert cognitive knowledge, causal reasoning knowledge, and facts relevant to the current prediction task—the core content framework of the task prompts is constructed. To guide the general generative language model to better understand and execute domain-specific prediction tasks, an example-based teaching method is employed. A small number of high-quality task examples, meticulously decomposed by humans, are prepared. These examples are integrated into the prompt template as demonstrations for model learning. In this way, the model can effectively learn from these limited demonstration examples the domain background knowledge, analytical paradigms, and output format requirements closely related to the task type. The prompt templates for event trend prediction tasks typically explicitly include teaching information such as the task type definition, task objective description, and specific output requirements.

[0047] Based on the event evolution reasoning prediction results and the results of related information vector retrieval, the steps for conflict detection of event elements through closed-loop evidence chain analysis include: Vector similarity retrieval technology is used to retrieve recent key information related to the prediction problem from background information as the associated information vector retrieval results; Event elements are extracted from related information, and conflict detection is performed on the event evolution reasoning prediction results based on content detection rules from three aspects: spatiotemporal conflict, target attribute conflict, and content logic conflict.

[0048] The system performs vector retrieval of related information, employing vector similarity retrieval technology. It transforms the prediction question and the content in the vast background information database into vector representations, calculates the similarity between vectors, and leverages an optimized index structure to quickly and accurately retrieve the most relevant recent key information from the background information that is most relevant to the prediction question.

[0049] Conflict detection calculates the overall conflict score using a weighted summation method. ,in Due to a conflict in time and space, Due to a conflict in target attributes, Due to logical conflicts in the content, All are weighting coefficients.

[0050] Event element conflict detection is performed by extracting structured event elements such as time, location, people, cause, purpose, and result from the retrieved related information. These elements are then compared with the preliminary predictions generated in the "Event Open Prediction" step. Conflict detection is conducted in three dimensions based on preset content detection rules: firstly, spatiotemporal conflicts. First, check whether there are contradictions between the prediction conclusions and related information in terms of time and location; second, check for conflicts in target attributes. First, check whether the attributes or goals of the individuals, organizations, and other entities involved are consistent; second, check for logical conflicts in the content. The process involves checking whether the causal, sequential, and other logical relationships of the events are reasonable. Finally, a weighted summation method is used to calculate the overall conflict score. ,in These are the weighting coefficients for the three conflict detections. This score is used to quantify the degree of inconsistency between the predicted conclusions and the known related information.

[0051] The process of using a false detection model to filter credible conclusions and obtain credible prediction conclusions includes: A false alarm template is constructed by linking information and inference predictions. The inference predictions are then filtered using a false alarm detection model, retaining those that pass the false alarm detection, and the inference predictions are ranked by their credibility. ,in For the false detection model to correlate information The detection rules As weight.

[0052] By integrating the obtained correlation information with the inference predictions to be verified, a dedicated false alarm detection prompt template is constructed. This template aims to guide the false alarm detection model in judging the consistency between the predicted conclusions and the factual basis. Then, this template is input into the pre-trained false alarm detection model, which will judge the authenticity of each inference prediction conclusion, filter out the conclusions judged as false or unreliable, and only retain the prediction results that pass the detection.

[0053] For multiple predictions that pass the detection, the system will further calculate a quantitative credibility score for them. And sort them. Credibility score The calculation is performed using a weighted summation method, and the specific formula is as follows: Ultimately, the system will score the reliability based on the calculated reliability rating. All retained conclusions are sorted to output a list of predictions ranked from highest to lowest confidence.

[0054] The method also includes: The system sorts the multiple predictions that pass the screening by credibility and outputs a list of the sorted credible predictions.

[0055] Following the false positive detection step, all the selected predictions are ranked by credibility. The core method involves calculating a comprehensive credibility score for each prediction, which is essentially a weighted sum of the detection results from various related information sources. Finally, the system outputs a ranked list of credible predictions based on their scores, from highest to lowest.

[0056] Figure 2This diagram illustrates another flowchart of the event evolution reasoning prediction method proposed in this application. Its core process first includes a knowledge construction and guided generation stage. This stage begins with cognitive logic learning, where structured cognitive logic knowledge is formed by extracting expert thinking and causal reasoning from domain-specific materials in parallel. Subsequently, the system proceeds to open event prediction, using this cognitive knowledge to understand the specific prediction task and constructing a prompt template that integrates expert thinking, causal reasoning, and problem facts, thereby guiding the large language model to generate preliminary intelligent guesses.

[0057] External verification and credible output retrieve relevant background information from the fact database through association information vector retrieval. Next, the preliminary conclusions are compared with the retrieved association information, undergoing multi-level verification and filtering through event element conflict detection and consistency association analysis. Finally, the system outputs verified and credible event evolution prediction conclusions, thus completing a full closed loop from knowledge guidance to evidence verification.

[0058] Compared with the prior art, the embodiments of this application have the following beneficial effects: First, traditional event evolution prediction techniques rely on open-ended model generation, making conclusions susceptible to model "illusion" and thus unstable. This invention, by constructing a cognitive logic knowledge prompt template that integrates expert thinking, causal reasoning, and problem facts, guides the model generation process into a structured analytical framework, thereby achieving a shift from "open-ended generation" to "template-assisted generation," significantly improving the consistency and reliability of the output conclusions.

[0059] Secondly, traditional methods lack effective verification mechanisms, making it difficult to guarantee the credibility of the conclusions generated. This invention employs a multi-level screening process, including associated information vector retrieval, rule-based event element conflict detection, and a false detection model, to cross-validate and verify the initially generated conclusions, forming a complete closed loop for credible conclusion screening. This significantly improves the credibility of the final output prediction conclusions.

[0060] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for predicting event evolution based on cognitively driven intelligent generation, characterized in that, include: By using a domain-wide language model, expert knowledge points, analysis processes, analytical perspectives, and logical reasoning are extracted from event research reports to construct an interactive thinking chain. Based on the event research reports, a domain event knowledge base is built to extract causal reasoning and construct a causal reasoning network structure. Based on the interactive thinking chain and causal reasoning network structure, cognitive logic knowledge and problem-related facts of the event to be predicted are integrated and embedded into the prompt template to construct a prompt model. This enables the domain-wide language model to generate event evolution reasoning prediction results based on expert analysis dimensions and analysis processes. Based on the event evolution reasoning prediction results and the related information vector retrieval results, conflict detection is performed on event elements through evidence chain closed-loop analysis, and credible conclusions are screened through a false detection model to obtain credible prediction conclusions.

2. The event evolution reasoning and prediction method as described in claim 1, characterized in that, The steps for extracting expert knowledge points, analysis processes, analytical perspectives, and logical reasoning from event research reports using a domain-specific large language model to construct an interactive thought chain include: Sentences from the event investigation report text Assign chapter element tags The elements of a text include an introduction, a central argument, sub-arguments, factual evidence, theoretical evidence, and a conclusion. Based on the identified text elements, structured extraction is performed to build an interactive thought chain, outputting problem analysis perspectives, analysis processes, and knowledge points.

3. The event evolution reasoning and prediction method as described in claim 1, characterized in that, The steps involved in building a domain event knowledge base based on event research reports, extracting causal principles, and constructing a causal principle network structure include: A method combining knowledge enhancement and prompting learning is adopted. A domain event knowledge base is built based on event research reports. Relevant domain event knowledge is injected into prompt templates and input into a general large language model to generate structured causal reasoning extraction results. Based on the results of causal reasoning extraction, a causal reasoning network structure is constructed. ,in For a set of event nodes, It is the set of edges representing causal relationships between events.

4. The event evolution reasoning and prediction method as described in claim 1, characterized in that, Based on an interactive thinking chain and causal network structure, the steps for constructing a prompt model by integrating cognitive logical knowledge with problem-related facts of the event to be predicted into a prompt template include: Based on cognitive knowledge, the task of predicting the development trend of events is analyzed and understood, the task type is identified, and key elements of the event background in the task text are extracted. The reasoning and prediction steps for constructing corresponding tasks based on interactive thinking chains ; Based on the task description instructions, inference and prediction steps, input task question sentences, and output format requirements, a prompt template is constructed. ,in For the general construction of prompt templates, For task description instructions, For task reasoning prediction steps , To input the task question sentence, For output format requirements.

5. The event evolution reasoning and prediction method as described in claim 4, characterized in that, The steps before constructing the prompt template include: Based on cognitive knowledge, the task of predicting the development trend of events is analyzed and understood, the task type is identified, and key elements of the event background in the task text are extracted.

6. The event evolution reasoning and prediction method as described in claim 1, characterized in that, When constructing the prompt template, it also includes: The generative language model is taught using manually decomposed examples so that it can learn task-related domain contextual knowledge from the examples.

7. The event evolution reasoning and prediction method as described in claim 1, characterized in that, Based on the event evolution reasoning prediction results and the results of related information vector retrieval, the steps for conflict detection of event elements through closed-loop evidence chain analysis include: Vector similarity retrieval technology is used to retrieve recent key information related to the prediction problem from background information as the associated information vector retrieval results; Event elements are extracted from related information, and conflict detection is performed on the event evolution reasoning prediction results based on content detection rules from three aspects: spatiotemporal conflict, target attribute conflict, and content logic conflict.

8. The event evolution reasoning and prediction method as described in claim 7, characterized in that, Conflict detection calculates the overall conflict score using a weighted summation method. ,in Due to a conflict in time and space, Due to a conflict in target attributes, Due to logical conflicts in the content, All are weighting coefficients.

9. The event evolution reasoning and prediction method as described in claim 1, characterized in that, The process of using a false detection model to filter credible conclusions and obtain credible prediction conclusions includes: A false alarm template is constructed by linking information and inference predictions. The inference predictions are then filtered using a false alarm detection model, retaining those that pass the false alarm detection, and the inference predictions are ranked by their credibility. ,in For the false detection model to correlate information The detection rules For weights.

10. The event evolution reasoning and prediction method as described in claim 1, characterized in that, The method further includes: The system sorts the multiple predictions that pass the screening by credibility and outputs a list of the sorted credible predictions.

Images