A multi-causal relationship extraction method and system for intelligent evaluation

Abstract

The present application belongs to the field of language processing, and specifically relates to a multi-causal relationship extraction method and system for intelligent evaluation, aiming to solve the problems of performance bottleneck, significant "bucket effect", and insufficient adaptability and integration caused by relying on multiple independent models. The present application comprises: constructing a causal analysis prompt template for generating step-by-step reasoning as a thinking chain prompt word, combining the original corpus to generate a fine-tuning corpus, inputting into a large language model for training, and adopting low-rank adaptation technology to fine-tune the large language model during the training process to obtain a causal relationship extraction model; inputting the corpus to be extracted into the causal relationship extraction model, and outputting the causal relationship extraction result; generating a comprehensive evaluation result and decision suggestion. The present application improves the efficiency and quality of the multi-causal relationship extraction task, enhances the stability and adaptability of the system, and at the same time reduces the labor cost and provides scientific and reliable evaluation results and decision suggestions.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of language processing, and particularly relates to a multi-causal relationship extraction method and system for intelligent evaluation. BACKGROUND

[0002] In the current technical framework, solutions for handling complex causal relationships often rely on a series of step-by-step methods, including but not limited to event element extraction, causal relationship identification, and causal type classification. However, these methods often rely on multiple independent models to complete their respective tasks, which leads to several key problems:

[0003] Performance bottleneck: Due to the uneven capabilities and accuracies of individual models, the overall performance of the system is often limited by the weakest link, the so-called "bottleneck effect". This phenomenon greatly affects the overall task performance.

[0004] Lack of adaptability: Especially in scenarios such as space domain with huge amounts of data and complex scenarios, traditional rule-based or single-model-based methods are not up to the task. They are difficult to efficiently handle large-scale data sets, and lack the necessary flexibility and adaptability when faced with different types of causal relationships.

[0005] Integration challenge: Integrating multiple independent models into an effective system is also a challenge. The interface design between different models, data format conversion, and other issues require additional development and maintenance costs, increasing the complexity and potential failure points of the system.

[0006] Based on this, the present application proposes a multi-causal relationship extraction method and system for intelligent evaluation. SUMMARY

[0007] To solve the above problems in the prior art, i.e. when handling complex causal relationships, due to the reliance on multiple independent models, performance bottlenecks and "bottleneck effects" are significant, and when faced with large-scale data, the adaptability and integration are insufficient, thus affecting the overall task performance and efficiency, the present application provides a multi-causal relationship extraction method and system for intelligent evaluation.

[0008] In a first aspect of the present application, a multi-causal relationship extraction method for intelligent evaluation is provided, which comprises:

[0009] Constructing a causal analysis prompt template for generating step-by-step reasoning as a thinking chain prompt word, and generating a fine-tuning corpus in combination with the original corpus; wherein the thinking chain prompt word contains multi-stage guidance instructions for text preprocessing, element extraction, and relationship determination;

[0010] inputting the fine-tuning corpus as input data into a large language model, training for a causal relationship, adopting a low-rank adaptation technology to fine-tune the large language model during the training process, and obtaining a causal relationship extraction model;

[0011] inputting a corpus to be extracted for a causal relationship into the causal relationship extraction model, and outputting a causal relationship extraction result;

[0012] Based on the causal relationship extraction result, combining a pre-constructed domain event evaluation index system and a constructed strategy knowledge base, a comprehensive evaluation result and a decision suggestion are generated.

[0013] Further, the thought chain prompt word, the construction process is:

[0014] Step A1, performing sentence processing and word segmentation analysis on the original corpus, identifying causal marker words in the sentence, and obtaining a set of independent sentences with causal markers;

[0015] Step A2, positioning the causal chain node based on the causal marker word, and extracting the core elements of the reason part and the result part, respectively, the core elements including action subject, action, action object, and space-time characteristics;

[0016] Step A3, determining whether the causal relationship belongs to direct causal relationship or indirect causal relationship containing intermediate events by analyzing the association path of the causal chain node;

[0017] Step A4, classifying the reason part and the result part into pre-set event categories, respectively, the event categories including at least one of military action, diplomatic activity, security event, political event, social event, technological development, economic event, aerospace activity, and equipment and arms;

[0018] Step A5, generating structured data in a predetermined format according to the core elements, the causal relationship type and the event category, the format including a causal relationship type field, a reason element set field and a result element set field;

[0019] Step A6, performing integrity check on the structured data through a multi-round verification mechanism, and dynamically adjusting the word segmentation rules and event classification model parameters based on the text features, to obtain the thought chain prompt word.

[0020] Further, in the training process, the low-rank adaptation technology is adopted to fine-tune the causal relationship extraction model, and the method is:

[0021] Step B1, dividing the fine-tuning corpus into a training set, a validation set and a test set;

[0022] Step B2, load a pre-trained large language model as a base model of the causal relationship extraction model, inject a low-rank adaptation matrix in the neural network layer of the base model to form a parameter subspace that can be fine-tuned;

[0023] Step B3, perform multi-round iterative training on the base model injected with the low-rank adaptation matrix through the training set, and perform hyperparameter optimization through the validation set to obtain a task-oriented model adapted to multi-causal relationship extraction;

[0024] Step B4, evaluate the model performance using the test set, and dynamically adjust the rank parameter of the low-rank matrix and the prompt word embedding strategy according to the evaluation results.

[0025] Further, the comprehensive evaluation result is obtained by:

[0026] A domain event evaluation index system is constructed and knowledge vectorization processing is performed; wherein the index system at least includes threat level, impact range and risk factor, and the index system is converted into vector form by knowledge embedding technology and stored in the evaluation knowledge base;

[0027] According to the causal relationship extraction result, the matched evaluation index is dynamically retrieved in the evaluation knowledge base through semantic similarity calculation, and an event special evaluation system is constructed;

[0028] A complex reasoning large model is taken as a base, the causal relationship extraction result and the event special evaluation system are fused for multi-level reasoning to generate a comprehensive evaluation result including risk factors, impact range and priority.

[0029] Further, the domain event evaluation index system is constructed by:

[0030] Public standard files, authoritative materials and industry research results of domain events are obtained, qualitative and quantitative evaluation standards are extracted, and the domain event evaluation index system is obtained.

[0031] Further, the decision suggestion is obtained by:

[0032] A strategy knowledge base is constructed, including collecting a strategy method set related to the domain event, the strategy method set including solutions, coping strategies and optimization solutions; the strategy method set is subjected to knowledge vectorization processing, converted into vectorized data that can be called by the large model and stored, forming a structured and searchable strategy knowledge base;

[0033] The comprehensive evaluation result is taken as input, the strategy knowledge base is dynamically retrieved by a knowledge base retriever, and the strategy information associated with the current event is extracted by semantic matching;

[0034] Based on a complex reasoning large model, the comprehensive evaluation result and the extracted strategy information are integrated to generate decision suggestions containing specific action plans, priorities and timeliness, and the feasibility and applicability of the suggestions are optimized through multiple rounds of reasoning, and the suggestions are output in the form of a chart or a text report.

[0035] Further, the specific process of generating decision suggestions is as follows:

[0036] According to the event background, risk factors and strategy information, a short-term action plan and a long-term optimization scheme are automatically generated;

[0037] The applicability of the suggestions is optimized through context adjustment to ensure that the risk level matches the evaluation result.

[0038] Further, the multiple rounds of reasoning process includes iterative verification of strategy feasibility, resource consumption and execution timeliness.

[0039] Further, the priority arrangement of the decision suggestions is based on the comprehensive determination of threat urgency, strategy implementation cost and risk mitigation effect.

[0040] The second aspect of the application provides a multi-causal relationship extraction system for intelligent evaluation, based on a multi-causal relationship extraction method for intelligent evaluation, the system comprises:

[0041] A database construction module configured to construct a cause-effect analysis prompt template for generating step-by-step reasoning as a thinking chain prompt word, and generate a fine-tuning corpus in combination with the original corpus; wherein the thinking chain prompt word contains multi-stage guidance instructions for text preprocessing, element extraction and relationship determination;

[0042] A training module configured to input the fine-tuning corpus as input data into a large language model, train the cause-effect relationship, and fine-tune the large language model using low-rank adaptation technology during the training process to obtain a cause-effect relationship extraction model;

[0043] A cause-effect relationship extraction module configured to input the corpus to be extracted into the cause-effect relationship extraction model, and output the cause-effect relationship extraction result;

[0044] An evaluation and suggestion module configured to generate a comprehensive evaluation result and a decision suggestion based on the cause-effect relationship extraction result, in combination with a pre-constructed domain event evaluation index system and a constructed strategy knowledge base.

[0045] The beneficial effects of the application are as follows:

[0046] Improving evaluation efficiency and quality: By using large models and their fine-tuning techniques, especially the LoRA technique, this method can efficiently handle the multi-causal relationship extraction task in domain events. This not only reduces human involvement and the probability of human error, but also ensures the transparency and interpretability of the evaluation results.

[0047] Large model outputs complete evaluation reasoning process, making each step clear and visible, thus improving the credibility of the evaluation results.

[0048] Reducing the "bottleneck effect" and illusion problem: The thinking chain instruction combined with the LoRA fine-tuning technique significantly reduces the illusion problem of large models when dealing with complex causal relationships and optimizes their instruction following ability. This makes the entire system more stable and reliable when facing massive data, further improving the accuracy and reliability of task execution.

[0049] By constructing thinking chain prompts containing text preprocessing, element extraction, and relationship judgment guiding instructions, the problem of single link becoming a performance bottleneck is effectively avoided, enhancing the overall performance of the system.

[0050] Enhancing adaptability and flexibility: Using RAG technology combined with vector database storage of index system, adopting three-dimensional coordinate evaluation method to automatically retrieve relevant index system, deeply analyzing causal relationships, completing risk level evaluation of domain events. This combination allows the system to flexibly respond to different types of data and task requirements.

[0051] The introduction of the knowledge base provides reliable decision support for users, enhancing the practicality and user-friendliness of the system.

[0052] Reducing labor costs: The entire process is highly automated, reducing the dependence on professionals and greatly reducing labor costs. At the same time, through the application of intelligent evaluation methods, work efficiency is improved and project cycles are shortened.

[0053] Improving the scientificity of comprehensive evaluation results and the effectiveness of decision-making suggestions: Combined with pre-constructed domain event evaluation index system and strategy knowledge base, comprehensive evaluation results and decision-making suggestions are generated, ensuring the scientificity of evaluation results and the feasibility of decision-making suggestions, providing strong support for practical applications. BRIEF DESCRIPTION OF DRAWINGS

[0054] Other features, objects and advantages of the present application will become more apparent from the following detailed description of non-limiting embodiments made with reference to the accompanying drawings:

[0055] Figure 1 is a flowchart of a multi-causal relationship extraction method for intelligent evaluation of the present application;

[0056] Figure 2 is a flowchart for fine-tuning training of the causal relationship extraction model in the multi-causal relationship extraction method for intelligent evaluation of the application;

[0057] Figure 3 is a flowchart for intelligent evaluation in the multi-causal relationship extraction method for intelligent evaluation of the application;

[0058] Figure 4 is a strategy generation flowchart in the multi-causal relationship extraction method for intelligent evaluation of the application. DETAILED DESCRIPTION

[0059] The application will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific examples described herein are intended to explain the relevant application, but not to limit the application. In addition, it should be noted that only the parts related to the application are shown in the drawings for ease of description.

[0060] It should be noted that the embodiments and features in the embodiments in the application can be combined with each other without conflict. The application will be described in detail below with reference to the accompanying drawings and examples.

[0061] The application provides a multi-causal relationship extraction method for intelligent evaluation, which comprises:

[0062] A causal analysis prompt template for generating step-by-step reasoning is constructed as a thinking chain prompt word, and a fine-tuning corpus is generated in combination with original corpus; wherein the thinking chain prompt word contains multi-stage guidance instructions for text preprocessing, element extraction and relationship determination;

[0063] The fine-tuning corpus is input as input data into a large language model, and the large language model is trained for causal relationship. In the training process, the large language model is fine-tuned by using low-rank adaptation technology, and a causal relationship extraction model is obtained;

[0064] The corpus to be extracted for causal relationship is input into the causal relationship extraction model, and the causal relationship extraction result is output;

[0065] Based on the causal relationship extraction result, a comprehensive evaluation result and decision suggestion are generated by combining a pre-constructed domain event evaluation index system and a constructed strategy knowledge base.

[0066] The application extracts and generates structured causal relationship data by extracting the causal relationship of the domain event, providing high-quality and reliable data foundation support for subsequent intelligent evaluation. Then, the structured causal relationship is combined with the index system in the knowledge base, and intelligent evaluation is carried out based on the large model to generate comprehensive and accurate evaluation results.

[0067] On this basis, the model generates reasonable decision suggestions by searching the strategy knowledge base, thereby providing efficient and accurate auxiliary decision support for the user.

[0068] In order to more clearly describe the multi-causal relationship extraction method for intelligent evaluation, the following describes the multi-causal relationship extraction method for intelligent evaluation in combination with Figure 1 The embodiments of the present application are described in detail as follows:

[0069] Referring to Figure 1 and Figure 2 , a cause-effect analysis prompt template for generating step-by-step reasoning is constructed as a thinking chain prompt word, and a fine-tuning corpus is generated in combination with original corpus; wherein the thinking chain prompt word contains multi-stage guidance instructions for text preprocessing, element extraction and relationship determination;

[0070] In this embodiment, a step-by-step prompt word based on logical reasoning, i.e., a thinking chain prompt word, is designed to guide the model to understand and extract causal relationships in a step-by-step reasoning manner, thereby improving the accuracy and logic of causal relationship extraction. Collect original corpus related to domain events, such as scientific literature, accident reports, technical instructions and news reports, etc. These corpus need to cover single cause-effect, multi-cause-effect and complex cause-effect network of various scenes, to provide comprehensive and high-quality data support for building a cause-effect extraction model.

[0071] In the present application, the construction process of the thinking chain prompt word is as follows:

[0072] Step A1, sentence processing and word segmentation analysis are performed on the original corpus, the causal marker words in the sentence are identified, and a set of independent sentences with causal markers is obtained;

[0073] Specifically, sentence processing: according to punctuation marks, the input text is divided into independent sentences for sentence-by-sentence analysis of causal relationships.

[0074] Word segmentation analysis: word segmentation is performed on each sentence to identify verbs, nouns, prepositions, conjunctions and other key components to preliminarily determine the possible cause-effect structure.

[0075] Identify causal marker words: detect causal marker words in the text, such as "because", "lead to", "so" and the like, to indicate the existence of causal relationships.

[0076] Step A2, locate the cause-effect chain node based on the causal marker words, and extract the core elements of the cause part and the result part, respectively, the core elements including action subject, action, action object, time and space characteristics;

[0077] Specifically, determine the cause and the result: by analyzing the causal marker words, identify the "cause" and "result" parts in the cause-effect chain.

[0078] Disassemble the core elements:

[0079] Cause; Actor: Extracting "who" did what;

[0080] Action: Specific behavior; Object: Related object; Time: Time of occurrence; Location: Location of occurrence; Effect: Also extracting "who", "what", "object", "time", "location".

[0081] Step A3, by analyzing the associated path of the cause-effect chain node, determine whether the cause-effect relationship belongs to direct cause-effect relationship or indirect cause-effect relationship containing intermediate events;

[0082] In the present application, it is necessary to fully understand the multi-layer cause-effect chain contained in the input text, check the multi-layer cause-effect relationship between events, and understand the complex interactive effects.

[0083] Direct cause-effect relationship: Identify that one event directly causes another event. For example, event A directly leads to event B.

[0084] Indirect cause-effect relationship: Identify the cause-effect relationship formed through intermediate events or conditions. For example, event A affects event B through one or more intermediate events (event C).

[0085] Step A4, classify the cause part and the result part into preset event categories respectively, the event categories including at least one of military action, diplomatic activity, security event, political event, social event, technological development, economic event, aerospace activity, and equipment and weaponry;

[0086] Step A5, according to the core elements, cause-effect relationship type and event category, generate structured data conforming to a predetermined format, the format including a cause-effect relationship type field, a cause element set field and a result element set field;

[0087] The extracted "cause" and "result" parts are output according to the preset structured format. The format needs to include the following information:

[0088] {

[0089] "causality_type" : "cause-effect relationship type",

[0090] "cause" : {

[0091] "actor" : "actor",

[0092] "class" : "event category",

[0093] "action" : "subject action",

[0094] "time" : "time information",

[0095] "location" : "location information",

[0096] "object" : "object"

[0097] },

[0098] "effect" : {

[0099] "actor" : "actor",

[0100] "class" : "event category",

[0101] "action" : "subject action",

[0102] "time" : "time information",

[0103] "location" : "location information",

[0104] "object" : "object".

[0105] Step A6: Integrity check of structured data through multi-round verification mechanism, and dynamic adjustment of word segmentation rules and event classification model parameters based on text features, to obtain thought chain prompt words.

[0106] Repeat the above steps for multiple sentences or paragraphs in the text to ensure that important causal chains are not missed.

[0107] Optimization adjustment: adjust word segmentation rules, event categories according to specific text characteristics to ensure that the output causal chain is complete and accurate.

[0108] The following is the output template:

[0109] {"causality_list" : [ {"causality_type" : "", "cause" : {"actor" : "", "class" : "", "action" : "", "time" : "", "location" : "", "object" : ""}, "effect" : {"actor" : "", "class" : "", "action" : "", "time" : "", "location" : "", "object" : ""}}]}.

[0110] Specifically, the logical and clear thought chain prompt words are added to the original corpus to guide the model to reason step by step, so as to more accurately understand and extract the cause-effect relationship. At the same time, the corpus needs to cover diversified cause-effect relationship modes, including direct cause-effect relationship, indirect cause-effect relationship and complex multi-cause-effect relationship network, to ensure the generalization ability of the model. In order to further improve the training quality, the cause-effect relationship in the corpus is manually annotated or semi-automatically annotated to generate a structured cause-effect relationship dataset. The annotation content includes cause-effect event pairs (such as "event A leads to event B") and the type and weight of the cause-effect relationship (such as direct and indirect), to provide detailed training basis.

[0111] The fine-tuning corpus is input as input data into the cause-effect relationship extraction model for training, and a low-rank adaptation technology is used to fine-tune the cause-effect relationship extraction model during the training process;

[0112] The corpus to be extracted for cause-effect relationship is input into the cause-effect relationship extraction model, and the cause-effect relationship extraction result is output;

[0113] Preferably, the cause-effect relationship extraction model in the embodiment is QWEN2.5-14B.

[0114] During the training process, the low-rank adaptation technology is used to fine-tune the cause-effect relationship extraction model, and the method is as follows:

[0115] Step B1, the fine-tuning corpus is divided into a training set, a validation set and a test set;

[0116] Step B2, load a pre-trained large language model as a base model of the cause-effect relationship extraction model, inject a low-rank adaptation matrix into the neural network layer of the base model to form a parameter subspace that can be fine-tuned;

[0117] Step B3, the base model injected with the low-rank adaptation matrix is trained by multiple rounds of iteration through the training set, and the hyperparameters are optimized in combination with the validation set to obtain a task-oriented model adapted to multi-cause-effect relationship extraction;

[0118] Step B4, evaluate the model performance by using the test set, and dynamically adjust the rank parameter of the low-rank matrix and the prompt word embedding strategy according to the evaluation result.

[0119] Specifically, in the construction of the causal relationship extraction model, the LoRA (Low Rank Adaptation) technique is used to fine-tune the base large model (Qwen2.5-14B). The optimization focuses on the reasoning module in the Qwen model. LoRA significantly reduces the training cost by introducing a low-rank matrix, while improving the model's adaptability to the causal relationship extraction task. Its advantage is that it can achieve efficient optimization under resource constraints without modifying the original model weights. In this process, first, the collected corpus is divided into training, validation, and test sets to build a high-quality fine-tuning corpus specifically for optimizing the causal relationship extraction task. We select the Qwen2.5 general pre-training model as the base model, input the labeled data in the fine-tuning corpus into the model, and perform multiple rounds of training and validation, and continuously adjust the model parameters to ensure that it can efficiently identify and extract causal relationships.

[0120] During training, the model's input not only includes the original corpus data, but also incorporates thought chain cues as auxiliary information to enhance the model's reasoning ability. Specifically, the thought chain cues are embedded in the "system prompt" part of each input sample, and then the input-output pairs before and after are extracted, corresponding to "user input" and "model output", respectively. This approach helps the model better capture the logic of causal relationships while understanding the original corpus.

[0121] After fine-tuning, a specialized large model for the domain event causal relationship extraction task is obtained. This model has the following characteristics:

[0122] (1) It can accurately extract causal relationships in text, including single causal, multiple causal, and complex network causal relationships.

[0123] (2) It outputs structured data (such as causal chains, causal graphs, etc.), providing standardized input for subsequent intelligent evaluation modules.

[0124] Application examples:

[0125] Input event text: "A country launches a satellite, leading to an increase in space debris, further affecting orbital resource competition."

[0126] Model output causal relationship:

[0127] {

[0128] "causality_list":[

[0129] {

[0130] "causality_type":"direct causal relationship",

[0131] "cause":{

[0132] "actor" : "A country",

[0133] "class" : "Aerospace activities",

[0134] "action" : "Launch",

[0135] "time" : "",

[0136] "location" : "",

[0137] "object" : "Satellite"

[0138] },

[0139] "effect" : {

[0140] "actor" : "",

[0141] "class" : "Aerospace activities",

[0142] "action" : "Increase",

[0143] "time" : "",

[0144] "location" : "",

[0145] "object" : "Space debris"},

[0146] {

[0147] "causality_type" : "Indirect causality",

[0148] "cause" : {

[0149] "actor" : "",

[0150] "class" : "Aerospace activities",

[0151] "action" : "Increase",

[0152] "time" : "",

[0153] "location" : "",

[0154] "object" : "Space debris"

[0155] },

[0156] "effect" : {

[0157] "actor" : "",

[0158] "class": "aerospace activities",

[0159] "action": "impact",

[0160] "time": "",

[0161] "location": "",

[0162] "object": "orbital resource contention"}]}.

[0163] Through the above process, the method constructs a domain-adapted causal relationship extraction large model, and realizes the whole process optimization from corpus preparation to model fine-tuning. The model can efficiently extract the causal relationship in the domain event, and provides a solid data foundation for intelligent evaluation and decision support.

[0164] Based on the causal relationship extraction result, combined with the pre-constructed domain event evaluation index system and the constructed strategy knowledge base, a comprehensive evaluation result and decision suggestion are generated.

[0165] The comprehensive evaluation result is obtained by:

[0166] A domain event evaluation index system is constructed and knowledge vectorization processing is performed; wherein the index system at least includes threat level, impact range and risk factor, and the index system is converted into vector form by knowledge embedding technology and stored in the evaluation knowledge base;

[0167] Based on the professional knowledge of the domain (taking space safety as an example), a comprehensive evaluation index system is first designed, which includes threat level, impact range, risk factor and other key indicators. Then, the designed index system is processed by knowledge vectorization, which is converted into a vectorized format that can be understood and called by the large model, and is stored in the evaluation knowledge base. Through knowledge vectorization, the model can quickly retrieve high-quality knowledge support, thereby ensuring the comprehensiveness and accuracy of the evaluation process.

[0168] In the process of constructing the domain event evaluation index system and performing knowledge vectorization processing, the evaluation index system is coded as a vector state representation, hierarchical similarity calculation is used to realize super-high-speed knowledge retrieval, and a hybrid retrieval architecture is developed to overlay vector fast retrieval association on the traditional knowledge graph.

[0169] Among them, the evaluation index system is coded as a vector state representation, which specifically includes:

[0170] C1, decompose the feature matrix of each evaluation index to extract the principal feature vector and its corresponding characteristic value;

[0171] Step C2, the approximate rank of the feature matrix is calculated by the matrix rank analysis module, and the number of storage fragments is dynamically allocated according to the formula Q = ceil (log2 (r)) + k, wherein r is the approximate rank, k is the redundancy coefficient based on the storage performance; in the embodiment, 0.5≤ k ≤2.

[0172] Step C3, the main feature vector is mapped to a multi-dimensional vector space by using a piecewise encoding technique, and the time effectiveness of the vector is modulated by time stamp weighting.

[0173] Step C4, the associated metadata label is generated to form an extensible vector representation unit.

[0174] The similarity calculation method specifically includes:

[0175] A family of hash functions is generated based on local sensitive hashing (LSH), and the number of hash buckets is dynamically set according to the size of the knowledge base M . The high-dimensional vector is mapped to a low-dimensional hash bucket.

[0176] A KD-Tree structure is constructed in each hash bucket, and a hyperplane is divided according to a feature variance weight, wherein the weight w j is the feature variance j / ∑feature variance.

[0177] The hash bucket to which the query vector belongs is calculated, and a candidate set C is extracted according to a candidate ratio γ, wherein gamma is negatively correlated with the number of hash buckets N .

[0178] The candidate set is screened in two stages:

[0179] First stage: the cosine similarity is calculated S cos , and the vectors satisfying S cos ≥0.7θ (t) are reserved.

[0180] Second stage:

[0181] The weighted Euclidean distance is calculated for the reserved vectors , and is converted into the final similarity S final =1- D w / max( D w ).

[0182] wherein, is the i-th jThe value of the i-th dimension of the j-th vector in the knowledge base. It represents the numerical value of the i-th dimension of the j-th vector in the knowledge base. For example, in the military event evaluation, if the evaluation index includes dimensions such as "threat level", "time sensitivity", etc. q j That is, the quantitative value of the user query in these specific dimensions.

[0183] v ij The value of the i-th dimension of the j-th vector in the knowledge base. It represents the numerical value of the i-th dimension of the j-th vector in the knowledge base. For example, in the military event evaluation, if the evaluation index includes dimensions such as "threat level", "time sensitivity", etc. i j

[0184] i The index of the vector in the knowledge base (such as the feature vector of the 35th historical event). i

[0185] The index of the feature dimension (such as the 7th feature value of the 35th event). j j

[0186] For example, if the knowledge base stores 1000 evaluation vectors of historical events, each event has 50 feature dimensions, then v 35,7 The value of the 7th feature of the 35th event (such as the "diplomatic impact factor").

[0187] The dual-engine index structure includes: a relational database layer and a vector engine layer;

[0188] is a database management system based on the relational model (table structure), such as MySQL, PostgreSQL, Oracle, etc. It stores data in the form of rows and columns, and performs data operations through SQL (Structured Query Language).

[0189] For example, the name of the event, the time, the classification label, the hierarchical relationship of the causal relationship chain, and other clear structured information.

[0190] Through SQL, it can quickly perform equality queries (such as WHERE event type ='military action'), range queries (such as time BETWEEN 2020 AND 2023), etc.

[0191] Ensure data consistency (such as ACID properties), suitable for operations that require high reliability (such as updating event status).

[0192] The vector engine layer is a storage and retrieval system specially designed for high-dimensional vector data, supporting approximate nearest neighbor search (ANN, Approximate Nearest Neighbor), such as Faiss, Milvus, Elasticsearch vector extension, etc.​​​​

[0193] The vector engine layer encodes unstructured data such as evaluation indicators, text semantics into vectors (such as 512-dimensional floating-point number arrays). It can quickly find the most similar vector set to the query vector (such as by cosine similarity or Euclidean distance).

[0194] For example, in military events, retrieve historical cases similar to the current battlefield situation.

[0195] Design a dynamic routing strategy. When the query complexity Qc exceeds the threshold Qt , start engine retrieval, otherwise use database retrieval,

[0196] where Qc = lambda 1 Causal chain depth + lambda 2 Number of time windows;

[0197] lambda 1 , lambda 2 Adjust the coefficient for the preset field.

[0198] Where, pre-load high-frequency query patterns to the memory cache area during system idle period, and dynamically maintain data activity through shard popularity analysis, cache refresh interval Δt=τ / log2(Ns),

[0199] tau is the data validity period, Ns is the number of active shards.

[0200] As Figure 3 shown, build an evaluation index system for field events, and the method is:

[0201] Obtain public standard files, authoritative materials and industry research results of field events, extract qualitative and quantitative evaluation standards, and use them as the field event evaluation index system.

[0202] According to the extraction results of causal relationships, dynamically retrieve matched evaluation indicators in the evaluation knowledge base through semantic similarity calculation, and build an event special evaluation system.

[0203] The causal relationship in the event is the basis for intelligent evaluation, which comes from the structured causal relationship data output by the fine-tuned causal relationship extraction large model, such as the causal chain. These data provide key basic information for intelligent evaluation, ensuring that the model can conduct in-depth analysis based on the causal relationship network and generate accurate evaluation results.

[0204] The evaluation knowledge base retrieval and matching is implemented by the knowledge base retriever. The system dynamically extracts relevant evaluation indicators from the knowledge base based on the input structured causal relationship data, and constructs an evaluation system that is most suitable for the current field event. For example, when the causal relationship involves "increase in space debris", the system automatically calls related indicators such as debris cleaning and orbit adjustment, ensuring the pertinence and accuracy of the evaluation.

[0205] Based on the complex reasoning large model, the causal relationship extraction results and the event special evaluation system are integrated for multi-level reasoning to generate comprehensive evaluation results including risk factors, impact range and priority.

[0206] The evaluation model is based on the complex reasoning large model (Deepseek-R1), combined with causal relationship data and the constructed evaluation index system for reasoning and comprehensive evaluation. The model analyzes the risk factors, impact range and priority of the event to generate a comprehensive evaluation result, and uses a clear and logical reasoning chain to ensure the interpretability of the evaluation result. To build this evaluation system, first, collect and analyze public standard documents, authoritative materials and industry research results for comprehensive summary. These materials provide the necessary theory and data support for the construction of evaluation standards. The specific process includes: collecting and analyzing standardized documents, industry manuals and academic research results related to the causal relationship extraction task to ensure the authority of the reference materials; refining qualitative evaluation standards such as the clarity of causal relationships and the rationality of reasoning paths, combining professional knowledge to ensure the comprehensiveness of the evaluation dimensions; establishing quantitative evaluation standards through historical data and cases, including impact degree, probability distribution and risk assessment, to ensure the objectivity and repeatability of the evaluation; combining qualitative and quantitative evaluation standards to form a comprehensive evaluation index system, supporting the accuracy and reliability of causal relationships, and providing risk management and decision support. The reasoning process uses a clear chain to ensure that the evaluation results are highly objective, scientific and interpretable.

[0207] After reasoning, the evaluation results of the field event can be obtained. The following is an example of a field event.

[0208] Event description: The increase in orbital debris directly threatens the safety of satellite operation, which in turn limits military exercises due to the inability to obtain reliable data.

[0209] Causal relationship:

[0210] {

[0211] "causality_list" : [

[0212] {

[0213] "causality_type" : "direct causal relationship",

[0214] "cause" : {

[0215] "actor" : "",

[0216] "class" : "Aerospace Activity",

[0217] "action" : "Increase",

[0218] "time" : "",

[0219] "location" : "Orbit",

[0220] "object" : "Debris"

[0221] },

[0222] "effect" : {

[0223] "actor" : "",

[0224] "class" : "Aerospace Activity",

[0225] "action" : "Threaten",

[0226] "time" : "",

[0227] "location" : "",

[0228] "object" : "Satellite Operational Safety"

[0229] }

[0230] },

[0231] {

[0232] "causality_type" : "Indirect Causality",

[0233] "cause" : {

[0234] "actor" : "",

[0235] "class" : "Aerospace Activity",

[0236] "action" : "Threaten",

[0237] "time" : "",

[0238] "location" : "",

[0239] "object" : "Satellite Operational Safety"

[0240] },

[0241] "effect" : {

[0242] "actor" : """",

[0243] "class" : "military operation",

[0244] "action" : "restricted",

[0245] "time" : """",

[0246] "location" : """",

[0247] "object" : "military exercise"}]}.

[0248] Output evaluation results:

[0249] Risk factors: Orbital debris (external environmental risk), satellite safety (technical risk), and restricted military exercises (strategic risk)

[0250] Impact range: Orbital debris directly threatens satellite safety, leading to restricted military capabilities.

[0251] Priority: High.

[0252] Through the above steps, the intelligent evaluation system realizes the full-process automation from the input of cause-effect relationship to the output of evaluation results. The system combines the cause-effect relationship network with the domain knowledge base, based on the reasoning ability of large models, to generate comprehensive, accurate and transparent evaluation results, providing a solid foundation for subsequent decision support.

[0253] As shown in Figure 4 , the decision suggestion is obtained by:

[0254] Building a strategy knowledge base, including collecting a set of strategy methods related to the domain event, the set of strategy methods including solutions, coping strategies and optimization schemes; knowledge vectorization processing is performed on the set of strategy methods, converting them into vectorized data that can be called by large models and storing them to form a structured and searchable strategy knowledge base;

[0255] The core of building a strategy knowledge base is to organize and collect a set of strategy methods related to the domain event, including solutions, coping strategies and optimization schemes, such as orbital debris removal technology, satellite orbit adjustment plan and debris monitoring mechanism. Through knowledge vectorization processing, these strategies are converted into vectorized data that can be understood and called by large models, and stored in the strategy knowledge base, thereby providing structured and searchable high-quality information for decision support, ensuring the scientificity and pertinence of decision suggestions.

[0256] The comprehensive evaluation results are inputted to retrieve the policy knowledge base dynamically through the knowledge base retriever, and the policy information associated with the current event is extracted by semantic matching;

[0257] The retrieval and matching of the policy knowledge base take the evaluation results (such as risk level, impact range, threat priority, etc.) generated by the intelligent evaluation module as input, and provide the decision-making module with a clear event background and core evaluation indicators, ensuring that the decision-making suggestions are highly relevant to the evaluation results. The knowledge base retriever dynamically retrieves the policy knowledge base according to the input content and extracts the most relevant policy information for the event. For example, when the evaluation result shows that "increased orbital debris threatens satellite safety", the retriever will preferentially call related policies for debris cleaning and orbital adjustment. Through semantic matching, the retriever extracts the best strategies from the knowledge base to solve the current problem, and these strategy information will be the key basis for generating decision-making suggestions.

[0258] Based on the complex reasoning large model (Deepseek-R1), the comprehensive evaluation results and extracted policy information are integrated to generate decision-making suggestions containing specific action plans, priorities, and time effectiveness, and through multiple rounds of reasoning, the feasibility and applicability of the suggestions are optimized, and the output is in the form of charts or text reports.

[0259] The specific process of generating decision-making suggestions is as follows:

[0260] Based on the event background, risk factors, and policy information, short-term action plans and long-term optimization schemes are automatically generated;

[0261] The applicability of the suggestions is optimized through context adjustment to ensure that the risk level matches the evaluation results.

[0262] The multiple rounds of reasoning process include iterative verification of the feasibility of the strategy, resource consumption, and execution timeliness.

[0263] The priority arrangement of the decision-making suggestions is based on the comprehensive judgment of threat urgency, policy implementation cost, and risk mitigation effect.

[0264] More specifically, the auxiliary decision generation is realized by the DeepSeek-R1 base model, which combines the evaluation results and the retrieved strategy information, analyzes the event background, risk factors and available strategies, and generates targeted decision suggestions. The model first integrates the input data to automatically generate specific action plans and priority arrangements, and optimizes the feasibility and applicability of the suggestions through multiple rounds of reasoning and context adjustment. For example, when the input evaluation result is "increased orbital debris → threatened satellite safety → limited military exercises; risk level: high", combined with the retrieved "debris cleaning technology" and "satellite orbit adjustment scheme", the model outputs decision suggestions including short-term actions: launch debris cleaning tasks, focusing on cleaning debris in the threat area; long-term plan: optimize satellite orbit design, establish real-time debris monitoring mechanism, reduce future risks. The final generated auxiliary decision suggestions include specific response plans, priorities and timeliness, and are presented intuitively in the form of charts or text reports, such as the recommended direction for "high-risk" events is to implement debris cleaning, adjust orbit design and establish monitoring mechanism to improve future response efficiency.

[0265] In summary, the auxiliary decision generation process based on intelligent evaluation, by combining evaluation results with strategy knowledge base, realizes the full-process automation from data analysis to decision suggestion under the support of the complex reasoning ability of DeepSeek-R1 large model. This method not only provides scientific and targeted decision suggestions, but also significantly improves the efficiency of handling complex tasks, providing strong support for strategic planning of complex tasks.

[0266] Although the steps in the above embodiments are described in the above order, those skilled in the art can understand that, in order to achieve the effect of the embodiments, the different steps do not have to be executed in such an order, and they can be executed simultaneously (in parallel) or in a reversed order, and these simple changes are within the protection scope of the present application.

[0267] The second embodiment of the present application is a multi-causal relationship extraction system for intelligent evaluation, based on a multi-causal relationship extraction method for intelligent evaluation. The system comprises:

[0268] A database construction module configured to construct a causal analysis prompt template for generating step-by-step reasoning as a thinking chain prompt word, and generate a fine-tuning corpus in combination with the original corpus; wherein the thinking chain prompt word contains multi-stage guidance instructions for text preprocessing, element extraction and relationship determination;

[0269] A training module configured to input the fine-tuning corpus as input data into a large language model, train for causal relationships, and fine-tune the large language model during the training process using low-rank adaptation technology to obtain a causal relationship extraction model;

[0270] a cause-effect relationship extraction module configured to input a corpus to be extracted for cause-effect relationship into a cause-effect relationship extraction model, and output a cause-effect relationship extraction result;

[0271] an evaluation and suggestion module configured to generate a comprehensive evaluation result and a decision suggestion based on the cause-effect relationship extraction result, in combination with a pre-constructed domain event evaluation index system and a constructed strategy knowledge base.

[0272] Those skilled in the art can clearly understand the specific working process of the system and the related description described above for the convenience and brevity of description, which can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0273] It should be noted that the multi-cause-effect relationship extraction system for intelligent evaluation provided in the foregoing embodiments is only exemplified by the division of the above functional modules, and in actual application, the above functions can be completed by different functional modules according to needs, that is, the modules or steps in the embodiments of the present application are further decomposed or combined, for example, the modules of the foregoing embodiments can be combined into one module, or can be further split into multiple sub-modules to complete all or part of the functions described above. The names of the modules and steps involved in the embodiments of the present application are only for distinguishing the modules and steps, and should not be considered as an improper limitation of the present application.

[0274] The electronic device of the third embodiment of the present application comprises:

[0275] at least one processor; and

[0276] a memory in communication connection with the at least one processor; wherein

[0277] The memory stores instructions executable by the processor, and the instructions are used to be executed by the processor to implement the multi-cause-effect relationship extraction method for intelligent evaluation.

[0278] The computer readable storage medium of the fourth embodiment of the present application stores computer instructions, and the computer instructions are used to be executed by the computer to implement the multi-cause-effect relationship extraction method for intelligent evaluation.

[0279] Those skilled in the art can clearly understand the specific working process of the storage device and the processing device and the related description described above for the convenience and brevity of description, which can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0280] Those skilled in the art should clearly understand that the modules and method steps of each example described in combination with the embodiments disclosed herein can be realized by electronic hardware, computer software or a combination of both. The programs corresponding to the software modules and method steps can be placed in a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art. In order to clearly illustrate the interchangeability of electronic hardware and software, the components and steps of each example have been described in the above description in general terms. Whether the functions are performed by electronic hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0281] The terms "first", "second", and the like are used to distinguish similar objects, rather than to describe or indicate a particular order or sequence.

[0282] The term "comprising" or any other similar term is intended to encompass non-exclusive inclusion, so that a process, method, article or device / apparatus including a series of elements includes not only those elements, but also other elements not explicitly listed, or inherent to the process, method, article or device / apparatus.

[0283] So far, the technical solutions of the present application have been described in combination with the preferred embodiments shown in the drawings, but those skilled in the art can easily understand that the protection scope of the present application is obviously not limited to these specific embodiments. Those skilled in the art can make equivalent changes or replacements to the related technical features without departing from the principles of the present application, and the technical solutions after such changes or replacements will fall within the protection scope of the present application.

Claims

1. A method for extracting multiple causal relationships for intelligent evaluation, characterized in that, The method includes: A causal analysis prompt template for generating step-by-step reasoning is constructed as a thought chain prompt word, and a fine-tuned corpus is generated by combining it with the original corpus; wherein, the thought chain prompt word includes multi-stage guidance instructions for text preprocessing, element extraction and relation determination; The construction process of the thought chain prompt words is as follows: the original corpus is processed by sentence segmentation and word segmentation analysis, causal marker words in the sentences are identified, and a set of independent sentences with causal markers is obtained; Based on the causal markers, the nodes of the causal chain are located, and the core elements of the cause and effect parts are extracted respectively. The core elements include the subject of action, the action, the object of action, and the spatiotemporal characteristics. By analyzing the association paths of the causal chain nodes, it can be determined whether the causal relationship is a direct causal relationship or an indirect causal relationship that includes intermediate events. The cause and effect parts are respectively classified into preset event categories, which include at least one of military operations, diplomatic activities, security incidents, political events, social events, technological developments, economic events, aerospace activities, and equipment and armaments. Based on the core elements, causal relationship types, and event categories, structured data conforming to a predetermined format is generated, which includes a causal relationship type field, a cause element set field, and a result element set field. The structured data is verified for integrity through a multi-round verification mechanism, and the word segmentation rules and event classification model parameters are dynamically adjusted based on text features to obtain thought chain prompt words; The fine-tuned corpus is used as input data into a large language model for training on causal relationships. During the training process, low-rank adaptation technology is used to fine-tune the large language model to obtain a causal relationship extraction model. Input the corpus from which causal relationships are to be extracted into the causal relationship extraction model and output the causal relationship extraction results; Based on the causal relationship extraction results, combined with the pre-constructed domain event evaluation index system and the construction strategy knowledge base, a comprehensive evaluation result and decision-making suggestions are generated. The comprehensive evaluation results are obtained using the following method: Construct a domain event assessment indicator system and perform knowledge vectorization processing; the indicator system should at least include threat level, scope of impact and risk factors, and the indicator system should be converted into vector form and stored in the assessment knowledge base through knowledge embedding technology; Based on the results of causal relationship extraction, matching evaluation indicators are dynamically retrieved from the evaluation knowledge base through semantic similarity calculation to construct an event-specific evaluation system; Based on a complex reasoning model, the system integrates the results of causal relationship extraction with the event-specific evaluation system to conduct multi-level reasoning and generate a comprehensive evaluation result that includes risk factors, scope of impact, and priority. The semantic similarity calculation includes: A family of hash functions is generated based on locality-sensitive hashing, and the number of hash buckets is dynamically set according to the size of the knowledge base, mapping high-dimensional vectors to low-dimensional hash buckets; A KD-Tree structure is constructed within each hash bucket, and the hyperplane is divided according to the feature variance weight, where the weight is the ratio of the feature variance to the total feature variance. Calculate the hash bucket to which the query vector belongs, and extract a candidate set according to the candidate ratio, wherein the candidate ratio is negatively correlated with the number of hash buckets; The candidate set is screened in two stages: the first stage calculates the cosine similarity and retains the vectors that meet the threshold condition; the second stage calculates the weighted Euclidean distance and converts it into the final similarity. A dual-engine index structure is adopted, including a relational database layer and a vector engine layer. A dynamic routing strategy is designed to start vector engine retrieval when the query complexity exceeds a threshold, otherwise database retrieval is used. The method for knowledge vectorization is as follows: For each evaluation indicator, perform feature matrix decomposition to extract the principal eigenvector and its corresponding eigenvalues; The approximate rank of the characteristic matrix is ​​calculated using the matrix rank analysis module, and the number of storage fragments is dynamically allocated according to the formula Q=ceil(log2(r))+k, where... r For approximate rank, k Redundancy coefficient based on storage performance; A segmented coding technique is used to map the main feature vector to a multi-dimensional vector space, and the timeliness of the vector is modulated by timestamp weighting. Generate associated metadata tags to form scalable vector representation units.

2. The method for extracting multiple causal relationships for intelligent evaluation according to claim 1, characterized in that, During training, a low-rank adaptation technique is used to fine-tune the causal relationship extraction model. The method is as follows: Step B1: Divide the fine-tuned corpus into a training set, a validation set, and a test set; Step B2: Load the pre-trained large language model as the base model of the causal relationship extraction model, and inject a low-rank adaptation matrix into the neural network layer of the base model to form a fine-tunable parameter subspace. Step B3: The base model with injected low-rank adaptation matrix is ​​trained through multiple rounds of iterative training using the training set, and hyperparameters are optimized by combining the validation set to obtain a task-oriented model adapted for multi-causal relationship extraction. Step B4: Use the test set to evaluate the model performance, and dynamically adjust the rank parameter of the low-rank matrix and the prompt word embedding strategy based on the evaluation results.

3. The method for extracting multiple causal relationships for intelligent evaluation according to claim 1, characterized in that, The method for constructing a domain event evaluation index system is as follows: Obtain publicly available standard documents, authoritative materials, and industry research findings on events in the field, and extract qualitative and quantitative evaluation standards to serve as an evaluation indicator system for events in the field.

4. The method for extracting multiple causal relationships for intelligent evaluation according to claim 1, characterized in that, The decision recommendations are obtained as follows: Constructing a strategy knowledge base includes collecting a set of strategy methods related to domain events, which includes solutions, coping strategies, and optimization schemes; performing knowledge vectorization processing on the strategy method set, transforming it into vectorized data that can be called by large models and storing it to form a structured and searchable strategy knowledge base; The comprehensive evaluation results are used as input, and the strategy knowledge base is dynamically retrieved through the knowledge base retrieval tool. Semantic matching is used to extract strategy information associated with the current event. Based on a complex reasoning model, the comprehensive evaluation results and extracted strategy information are integrated to generate decision recommendations that include specific action plans, priorities, and timeliness. The feasibility and applicability of the recommendations are optimized through multiple rounds of reasoning, and the results are output in the form of charts or text reports.

5. The method for extracting multiple causal relationships for intelligent evaluation according to claim 4, characterized in that, The specific process for generating decision recommendations is as follows: Based on the event background, risk factors, and strategy information, short-term action plans and long-term optimization solutions are automatically generated. By adjusting the context, the applicability of the recommendations is optimized to ensure they match the risk level of the assessment results.

6. The method for extracting multiple causal relationships for intelligent evaluation according to claim 4, characterized in that, The multi-round reasoning process includes iterative verification of strategy feasibility, resource consumption, and execution timeliness.

7. The method for extracting multiple causal relationships for intelligent evaluation according to claim 4, characterized in that, The prioritization of the decision recommendations is based on a comprehensive assessment of threat urgency, strategy implementation costs, and risk mitigation effectiveness.

8. A multi-causal relationship extraction system for intelligent evaluation, based on the multi-causal relationship extraction method for intelligent evaluation according to any one of claims 1-7, characterized in that, The system includes: The database construction module is configured to build causal analysis prompt templates for generating step-by-step reasoning as thought chain prompts, and combine them with the original corpus to generate a fine-tuned corpus; wherein, the thought chain prompts include multi-stage guidance instructions for text preprocessing, element extraction and relation determination; The construction process of the thought chain prompt words is as follows: the original corpus is processed by sentence segmentation and word segmentation analysis, causal marker words in the sentences are identified, and a set of independent sentences with causal markers is obtained; Based on the causal markers, the nodes of the causal chain are located, and the core elements of the cause and effect parts are extracted respectively. The core elements include the subject of action, the action, the object of action, and the spatiotemporal characteristics. By analyzing the association paths of the causal chain nodes, it can be determined whether the causal relationship is a direct causal relationship or an indirect causal relationship that includes intermediate events. The cause and effect parts are respectively classified into preset event categories, which include at least one of military operations, diplomatic activities, security incidents, political events, social events, technological developments, economic events, aerospace activities, and equipment and armaments. Based on the core elements, causal relationship types, and event categories, structured data conforming to a predetermined format is generated, which includes a causal relationship type field, a cause element set field, and a result element set field. The structured data is verified for integrity through a multi-round verification mechanism, and the word segmentation rules and event classification model parameters are dynamically adjusted based on text features to obtain thought chain prompt words; The training module is configured to input the fine-tuned corpus as input data into the large language model, train it for causal relationships, and use low-rank adaptation technology to fine-tune the large language model during the training process to obtain a causal relationship extraction model. The causal relationship extraction module is configured to input the corpus from which causal relationships are to be extracted into the causal relationship extraction model and output the causal relationship extraction results. The evaluation and recommendation module is configured to extract results based on the causal relationship, combine them with a pre-built domain event evaluation index system and a construction strategy knowledge base, and generate comprehensive evaluation results and decision recommendations. The comprehensive evaluation results are obtained using the following method: Construct a domain event assessment indicator system and perform knowledge vectorization processing; the indicator system should at least include threat level, scope of impact and risk factors, and the indicator system should be converted into vector form and stored in the assessment knowledge base through knowledge embedding technology; Based on the results of causal relationship extraction, matching evaluation indicators are dynamically retrieved from the evaluation knowledge base through semantic similarity calculation to construct an event-specific evaluation system; Based on a complex reasoning model, the system integrates the results of causal relationship extraction with the event-specific evaluation system to conduct multi-level reasoning and generate a comprehensive evaluation result that includes risk factors, scope of impact, and priority. The semantic similarity calculation includes: A family of hash functions is generated based on locality-sensitive hashing, and the number of hash buckets is dynamically set according to the size of the knowledge base, mapping high-dimensional vectors to low-dimensional hash buckets; A KD-Tree structure is constructed within each hash bucket, and the hyperplane is divided according to the feature variance weight, where the weight is the ratio of the feature variance to the total feature variance. Calculate the hash bucket to which the query vector belongs, and extract a candidate set according to the candidate ratio, wherein the candidate ratio is negatively correlated with the number of hash buckets; The candidate set is screened in two stages: the first stage calculates the cosine similarity and retains the vectors that meet the threshold condition; the second stage calculates the weighted Euclidean distance and converts it into the final similarity. A dual-engine index structure is adopted, including a relational database layer and a vector engine layer. A dynamic routing strategy is designed to start vector engine retrieval when the query complexity exceeds a threshold, otherwise database retrieval is used. The method for knowledge vectorization is as follows: For each evaluation indicator, perform feature matrix decomposition to extract the principal eigenvector and its corresponding eigenvalues; The approximate rank of the characteristic matrix is ​​calculated using the matrix rank analysis module, and the number of storage fragments is dynamically allocated according to the formula Q=ceil(log2(r))+k, where... r For approximate rank, k Redundancy coefficient based on storage performance; A segmented coding technique is used to map the main feature vector to a multi-dimensional vector space, and the timeliness of the vector is modulated by timestamp weighting. Generate associated metadata tags to form scalable vector representation units.

Images