Event analysis model training method and system based on progressive multi-view exploration

Abstract

The invention discloses an event analysis model training method and system based on progressive multi-view exploration, and the method comprises the steps: constructing a multi-view sample generation engine, and generating a multi-view answer set for the same event sample through a differential prompt template by using a heterogeneous knowledge base and a multi-source event corpus; performing semantic deduplication and consistency verification on the multi-view answer set, calculating semantic similarity between answers based on a pre-training language model, performing clustering deduplication on the answers with the similarity exceeding a threshold value, and verifying effectiveness and timeliness of reference information through a knowledge graph; establishing a difficulty quantitative evaluation matrix, performing objective and subjective fusion difficulty grading on the training samples, and calculating a comprehensive difficulty score by combining the two results; and dividing the training sample into a plurality of difficulty grades according to the comprehensive difficulty score, and executing progressive staged training based on a grading result, so that the model is gradually transited from a low-difficulty sample to a high-difficulty sample, and multi-view fusion learning and progressive ability optimization are completed.

Description

Technical Field

[0001] This invention relates to the field of natural language processing, and in particular to a training method and system for an event analysis model based on progressive multi-perspective exploration. Background Technology

[0002] With the widespread application of artificial intelligence technology in text understanding, semantic reasoning, and complex event analysis, improving the generalization ability and reasoning depth of models in multidimensional semantic tasks has become an important research direction for intelligent analysis systems. Most current mainstream event analysis models rely on end-to-end fine-tuning strategies and single prompt templates for training. While these models can achieve a certain level of accuracy on specific tasks, they still have significant shortcomings in scenarios such as multi-perspective information fusion, semantic ambiguity resolution, and complex causal chain reasoning.

[0003] On the one hand, existing models rely on a relatively simple method for constructing training samples, often depending on fixed standard answers or single-path supervision signals. This makes them prone to getting stuck in local optima when faced with inputs that are multifaceted or semantically ambiguous, hindering their ability to make comprehensive judgments based on multidimensional information. On the other hand, the training process generally lacks dynamic adjustment mechanisms based on sample difficulty, making it easy for models to experience training oscillations and performance degradation when directly transferring from simple to complex tasks. Furthermore, existing training paradigms fail to fully utilize the knowledge complementarity between multiple models, limiting the adaptability and innovation of models in open semantic environments.

[0004] In recent years, methods such as curriculum learning and multi-view learning have been gradually introduced into the model training process to mimic the gradual learning pattern of human cognition, which progresses from simple to complex. However, the integration of these methods in event-level semantic reasoning tasks is still in the initial exploratory stage, and a mature framework for difficulty quantification and dynamic fusion of multiple perspectives has not yet been formed. This results in significant bottlenecks in models regarding complex semantic associations, logical chain reasoning, and interpretability.

[0005] Therefore, there is an urgent need for a training method that can comprehensively utilize multi-source knowledge, construct multi-perspective samples, and perform progressive optimization based on sample difficulty, so as to achieve a leapfrog improvement in the event analysis model from element identification to deep reasoning ability, and enhance its stability and interpretability in complex semantic environments. Summary of the Invention

[0006] To address the aforementioned technical problems in the existing technology, this invention proposes a training method and system for an event analysis model based on progressive multi-perspective exploration.

[0007] According to a first aspect of the present invention, a training method for an event analysis model based on progressive multi-view exploration is proposed, comprising:

[0008] S1: Construct a multi-perspective sample generation engine, utilizing heterogeneous knowledge bases and multi-source event corpora, and generate a set of multi-perspective answers for the same event sample through differentiated prompt templates;

[0009] S2: Perform semantic deduplication and consistency verification on the multi-perspective answer set, calculate the semantic similarity between answers based on the pre-trained language model, perform clustering deduplication on answers with similarity exceeding the threshold, and verify the validity and timeliness of the cited information through the knowledge graph;

[0010] S3: Establish a difficulty quantification assessment matrix to classify the training samples using a combination of objective and subjective difficulty levels. The objective dimensions include semantic complexity, clause coverage, and inference chain complexity, while the subjective dimensions are determined based on model scores. The combined results of both are then used to calculate the overall difficulty score.

[0011] S4: Based on the comprehensive difficulty score, the training samples are divided into multiple difficulty levels, and progressive phased training is performed based on the classification results, so that the model gradually transitions from low-difficulty samples to high-difficulty samples, completing multi-view fusion learning and progressive optimization of capabilities.

[0012] In some specific embodiments, differentiated prompt templates include semantic parsing type, contextual analogy type, and element aggregation type templates.

[0013] In some specific embodiments, the multi-perspective sample generation engine calls multiple heterogeneous models, including a fine-tuned model for event analysis and a general large language model, and forms a multi-perspective answer set by comparing the semantic differences of the results generated by different models.

[0014] In some specific implementations, semantic deduplication uses the BERT-legal model to calculate the semantic similarity between answers, and clusters answers with similarity exceeding a threshold θ to remove duplicates. The threshold θ is set to 0.85.

[0015] In some specific embodiments, the overall score for the objective dimensions is determined according to the following formula: D k = 0.4a + 0.3b + 0.3c, where a represents semantic complexity, b represents clause coverage, and c represents reasoning chain complexity.

[0016] In some specific embodiments, the overall difficulty score is determined by the objective score D. k Subjective score D z The overall difficulty score was determined to be D. final =0.6D k +0.4(D z / 100).

[0017] In some specific embodiments, the progressive phased training in S4 specifically includes: in the first stage, training is performed based on low-difficulty samples to build the basic reasoning ability of the model; in the second stage, after the model has stabilized on low-difficulty samples, medium-difficulty samples are introduced for mixed training to enhance the ability; in the third stage, when the model's pass rate on medium-difficulty samples reaches a preset standard, the training ratio of high-difficulty samples is gradually increased based on the difficulty quantification evaluation results.

[0018] According to a second aspect of the invention, a computer-readable storage medium is provided on which one or more computer programs are stored, which, when executed by a computer processor, implement the method described above.

[0019] According to a third aspect of the present invention, an event analysis model training system based on progressive multi-view exploration is proposed, comprising:

[0020] The multi-perspective sample generation unit is configured to build a multi-perspective sample generation engine. It utilizes a heterogeneous knowledge base and multi-source event corpus to generate a set of multi-perspective answers for the same event sample through differentiated prompt templates.

[0021] The answer processing unit is configured to perform semantic deduplication and consistency verification on a multi-perspective answer set, calculate the semantic similarity between answers based on a pre-trained language model, perform clustering deduplication on answers with similarity exceeding a threshold, and verify the validity and timeliness of reference information through a knowledge graph.

[0022] The difficulty quantification assessment unit is configured to build a difficulty quantification assessment matrix, performing a combined objective and subjective difficulty classification on the training samples. The objective dimensions include semantic complexity, clause coverage, and inference chain complexity, while the subjective dimensions are determined based on model scores. The overall difficulty score is calculated by combining the results of both approaches.

[0023] The progressive training unit is configured to divide training samples into multiple difficulty levels based on the comprehensive difficulty score, and perform progressive phased training based on the classification results, so that the model gradually transitions from low-difficulty samples to high-difficulty samples, completing multi-view fusion learning and progressive optimization of capabilities.

[0024] In some specific embodiments, the differentiated prompt templates include semantic parsing type, contextual analogy type and element aggregation type templates; the multi-perspective sample generation engine calls multiple heterogeneous models, including event analysis domain fine-tuning models and general large language models, and forms a multi-perspective answer set by comparing the semantic differences of the results generated by different models.

[0025] In some specific implementations, semantic deduplication uses the BERT-legal model to calculate the semantic similarity between answers, and clusters answers with similarity exceeding a threshold θ to remove duplicates. The threshold θ is set to 0.85.

[0026] In some specific embodiments, the overall difficulty score is determined by the objective score D. k Subjective score D z The overall difficulty score was determined to be D. final =0.6D k +0.4(D z / 100), the overall score for the objective dimensions is determined according to the following formula: D k = 0.4a + 0.3b + 0.3c, where a represents semantic complexity, b represents clause coverage, and c represents reasoning chain complexity.

[0027] This invention proposes a training method and system for event analysis models based on progressive multi-perspective exploration. By constructing a multi-perspective sample generation engine, a difficulty quantification assessment module, and a progressive training mechanism, it achieves an overall improvement in the semantic understanding, reasoning depth, and training stability of the event analysis model. This invention utilizes multi-perspective answers generated by heterogeneous models, prompting the model to absorb knowledge from different semantic angles during training, significantly enhancing the semantic connections and logical reasoning abilities between event elements. By introducing a subjective-objective integrated difficulty quantification assessment system, training samples are progressively learned in layers according to difficulty levels, enabling the model to gradually master complex semantic structures along a path from easy to difficult, resulting in a more stable training process and improved convergence speed and accuracy. Simultaneously, based on a dynamic course scheduling mechanism, the model can adaptively adjust the sample ratio and weights after each training stage, effectively improving cross-scenario generalization performance and robustness. Furthermore, combined with a knowledge graph and semantic consistency verification mechanism, the reasoning chain output by the model possesses high interpretability and decision transparency. Overall, this invention significantly improves the comprehensive judgment accuracy, training efficiency, and interpretability of event analysis models. Attached Figure Description

[0028] The accompanying drawings are included to provide a further understanding of the embodiments and are incorporated in and constitute a part of this specification. The drawings illustrate embodiments and, together with the description, serve to explain the principles of the invention. Other embodiments and many anticipated advantages of the embodiments will be readily recognized as they become better understood through reference to the following detailed description. Other features, objects, and advantages of this application will become more apparent from reading the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0029] Figure 1 This is a flowchart of an event analysis model training method based on progressive multi-view exploration, according to an embodiment of this application.

[0030] Figure 2 This is a flowchart of a specific embodiment of the event analysis model training method based on progressive multi-view exploration in this application;

[0031] Figure 3 This is a flowchart of a progressive training strategy according to a specific embodiment of this application;

[0032] Figure 4 This is a flowchart illustrating the framework of a progressive multi-view exploration-based event analysis model training method, which is a specific embodiment of this application.

[0033] Figure 5 This is a schematic diagram of the structure of a computer system used to implement the electronic device of the present application. Detailed Implementation

[0034] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0035] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0036] Figure 1 A flowchart illustrating a method for training an event analysis model based on progressive multi-view exploration according to an embodiment of this application is shown. Figure 1 As shown, the method includes the following steps:

[0037] S1: Construct a multi-perspective sample generation engine, utilizing heterogeneous knowledge bases and multi-source event corpora, and generate a set of multi-perspective answers for the same event sample through differentiated prompt templates.

[0038] In specific embodiments, differentiated prompt templates include semantic parsing, contextual analogy, and element aggregation templates. The multi-perspective sample generation engine calls multiple heterogeneous models, including a fine-tuned model for event analysis and a general large language model, and forms a multi-perspective answer set by comparing the semantic differences between the results generated by different models.

[0039] S2: Perform semantic deduplication and consistency verification on the multi-perspective answer set, calculate the semantic similarity between answers based on the pre-trained language model, perform clustering deduplication on answers with similarity exceeding the threshold, and verify the validity and timeliness of the cited information through the knowledge graph.

[0040] In a specific implementation, semantic deduplication uses the BERT-legal model to calculate the semantic similarity between answers, and clusters answers with similarity exceeding a threshold θ to remove duplicates. The threshold θ is set to 0.85.

[0041] S3: Establish a difficulty quantification assessment matrix to classify the training samples using a combination of objective and subjective difficulty levels. The objective dimensions include semantic complexity, clause coverage, and inference chain complexity, while the subjective dimensions are determined based on model scores. The combined results of both are then used to calculate a comprehensive difficulty score.

[0042] In a specific embodiment, the overall score of the objective dimension is determined according to the following formula: D k = 0.4a + 0.3b + 0.3c, where a represents semantic complexity, b represents clause coverage, and c represents reasoning chain complexity. The overall difficulty score is determined by the objective score D. k Subjective score D z The overall difficulty score was determined to be D. final =0.6D k +0.4(D z / 100).

[0043] S4: Based on the comprehensive difficulty score, the training samples are divided into multiple difficulty levels, and progressive phased training is performed based on the classification results, so that the model gradually transitions from low-difficulty samples to high-difficulty samples, completing multi-view fusion learning and progressive optimization of capabilities.

[0044] In a specific embodiment, progressive phased training includes: in the first stage, training is conducted based on low-difficulty samples to build the basic reasoning ability of the model; in the second stage, after the model has stabilized on low-difficulty samples, medium-difficulty samples are introduced for mixed training to enhance the ability; in the third stage, when the model's pass rate on medium-difficulty samples reaches a preset standard, the training ratio of high-difficulty samples is gradually increased based on the difficulty quantification evaluation results.

[0045] Figure 2 A flowchart illustrating the framework of a progressive multi-view exploration-based event analysis model training method according to a specific embodiment of this application is shown, as follows: Figure 2 As shown, this method trains the model through three modules: a multi-view answer generation engine 201, a difficulty quantification evaluator 202, and a progressive training strategy 203. The specific implementation method is as follows:

[0046] The multi-perspective answer generation engine 201 generates high-quality multi-perspective training samples through heterogeneous knowledge fusion. This engine first integrates multi-source knowledge bases and semantic databases, and then generates a multi-dimensional set of answers for the same event sample based on differentiated prompt templates. Differentiated prompt templates include semantic parsing, contextual analogy, and element aggregation types to ensure that the generated answers are semantically differentiated and complementary. During the multi-perspective answer generation process, multiple heterogeneous AI models (including domain-specific fine-tuned models and general-purpose large language models) can be simultaneously invoked to obtain a multi-perspective answer set {A1, A2, ..., A...}. n}

[0047] To ensure the consistency and validity of the training samples, the multi-view answer generation engine 201 further performs standardization processing, including the following steps:

[0048] Semantic deduplication: Based on a BERT-like semantic embedding model, the semantic similarity between answers is calculated. When the similarity is higher than the threshold θ = 0.85, clustering is performed to eliminate redundant content.

[0049] Knowledge verification: The validity and timeliness of the knowledge entries or concept nodes cited in each answer are verified through a knowledge graph matching mechanism;

[0050] Conflict detection: When the degree of divergence Δ between multiple answers in terms of semantic stance or reasoning direction is greater than 0.4 (quantified by the stance classification model), a manual or model-level review mechanism is automatically triggered to ensure the reliability of the training data.

[0051] The difficulty quantification evaluator 202 is used to perform a subjective and objective difficulty assessment of samples and generate a three-dimensional quantitative grading result. In the objective dimension, three core evaluation indicators are set: semantic complexity, knowledge coverage, and reasoning chain complexity, corresponding to weight coefficients of 0.4, 0.3, and 0.3, respectively. Its comprehensive difficulty score can be expressed as: D k =0.4a+0.3b+0.3c, where a is the semantic complexity index of the embedding vector variance measure, b is the product of the number of knowledge references and the hierarchical coefficient, and c is the weighted value of the logical connective count and the causal graph depth.

[0052] On the subjective dimension, the training samples are scored using the base model and divided into simple samples (D) according to percentiles. z ≥80 points), medium sample (60≤D) z <80) and difficult samples (D z <60). Then, considering the objective difficulty level D... k Subjective score D z Calculate the final difficulty label: D final =0.6D k +0.4(D z / 100) to achieve adaptive stratification of samples.

[0053] The progressive training strategy module 203 divides samples into training sets of different stages based on difficulty labels, and performs phased learning in a progressive manner from low to high difficulty. The model first builds basic reasoning capabilities on low-difficulty samples. Once training is stable, medium- and high-difficulty samples are introduced for mixed training to achieve multi-view semantic fusion and capability enhancement. Specifically, Figure 3 A flowchart illustrating a progressive training strategy of a specific embodiment of this application is shown, such as... Figure 3 As shown, the progressive training strategy module of this embodiment adopts a three-stage course learning mechanism, which includes a basic ability building stage 310, a progressive enhancement stage 320, and a multi-view fusion stage 330, wherein each stage corresponds to training samples of different difficulty levels: simple samples 311, medium samples 321, and difficult samples 331. The final training results are used to improve the comprehensive ability of the event analysis model 204.

[0054] In a specific embodiment, during the first stage of basic capability building (310), the system selects samples (311) marked as low difficulty after difficulty quantification assessment as the training set and performs preliminary training to establish the model's basic semantic understanding and reasoning capabilities. By conducting stable low-difficulty learning in this stage, the risk of gradient fluctuations during the initial training process of the model can be effectively reduced, and a foundation can be laid for subsequent complex tasks.

[0055] In a specific embodiment, during the progressive enhancement phase 320 of the second stage, when the model converges and exhibits high stability on low-difficulty samples, the training system automatically introduces medium-difficulty samples 321 for mixed training. This stage aims to enhance the model's semantic generalization and feature transfer capabilities, enabling it to maintain high judgment accuracy in more complex event descriptions and multi-causal semantic relationships.

[0056] In a specific embodiment, during the third stage of multi-perspective fusion 330, the system gradually increases the training proportion of difficult samples 331 based on the feedback results from the difficulty quantification evaluator, and introduces diverse samples generated by models from different perspectives for joint training. This stage, through multi-dimensional feature fusion and dynamic weight optimization mechanisms, enables the model to simultaneously learn reasoning patterns from different semantic perspectives in highly complex scenarios, achieving cross-sample and cross-domain knowledge fusion and capability progression.

[0057] Through the three-stage progressive learning mechanism described above, the event analysis model of this invention can be systematically trained according to the learning path of "from easy to difficult, from single to multiple perspectives", thereby significantly improving the model's comprehensive semantic understanding ability, logical reasoning depth and cross-scene adaptability while maintaining training stability.

[0058] Ultimately, after completing the three-stage training process described above, the event analysis model 204 acquires multi-dimensional reasoning capabilities with high robustness and interpretability, outputting structured analysis results for tasks such as complex event semantic understanding, risk identification, and intelligent decision-making.

[0059] This invention proposes a training method for an event analysis model based on progressive multi-perspective exploration. By constructing a multi-perspective answer generation engine, a dynamic difficulty assessment mechanism, and a phased progressive training framework, it achieves a comprehensive improvement in the model's performance, stability, and interpretability. This method significantly enhances the model's comprehensive judgment ability through multi-perspective sample fusion, improving the event analysis accuracy by approximately 15% compared to traditional methods. The progressive training strategy, from easy to difficult, effectively improves the model's convergence speed by approximately 20% and reduces computational resource consumption. Combining knowledge graphs and semantic reasoning path annotations enables the model's output to possess high transparency and traceability, significantly enhancing decision interpretability. This model can be widely applied to scenarios such as compliance review, risk identification, public safety assessment, and intelligent semantic analysis, providing efficient and reliable technical support for automated decision-making in complex semantic environments, and possesses outstanding innovative value and promising application prospects.

[0060] Figure 4 This application illustrates an embodiment of an event analysis model training system architecture based on progressive multi-view exploration, as shown in the diagram. Figure 4 As shown, the system includes a multi-view sample generation unit 401 and an answer processing unit.

[0061] 402, Difficulty Quantification Assessment Unit 403, and Progressive Training Unit 404, including the multi-view sample generation unit.

[0062] Configuration 401 is used to build a multi-perspective sample generation engine, utilizing heterogeneous knowledge bases and multi-source event corpora to generate a set of multi-perspective answers for the same event sample through differentiated prompt templates; Configuration 402 is used to perform semantic deduplication and consistency verification on the multi-perspective answer set, calculate the semantic similarity between answers based on a pre-trained language model, perform clustering deduplication on answers with similarity exceeding a threshold, and verify the validity and timeliness of cited information through a knowledge graph; Configuration 403 is used to establish a difficulty quantification evaluation matrix, perform objective and subjective difficulty grading on training samples, where the objective dimension includes semantic complexity, clause coverage, and inference chain complexity, and the subjective dimension is determined based on model scoring, and the comprehensive difficulty score is calculated by combining the results of both; Configuration 404 is used to divide training samples into multiple difficulty levels according to the comprehensive difficulty score, and perform progressive phased training based on the grading results, so that the model gradually transitions from low-difficulty samples to high-difficulty samples, completing multi-perspective fusion learning and progressive optimization of capabilities.

[0063] The following is for reference. Figure 5 It shows a schematic diagram of the structure of a computer system suitable for implementing the electronic device of the present application. Figure 5 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0064] like Figure 5 As shown, the computer system includes a central processing unit (CPU) 501, which can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) 502 or programs loaded from storage section 508 into random access memory (RAM) 503. RAM 503 also stores various programs and data required for the operation of system 500. CPU 501, ROM 502, and RAM 503 are interconnected via bus 504. Input / output (I / O) interface 505 is also connected to bus 504.

[0065] The following components are connected to I / O interface 505: an input section 506 including a keyboard, mouse, etc.; an output section 507 including a liquid crystal display (LCD) and speakers, etc.; a storage section 508 including a hard disk, etc.; and a communication section 509 including a network interface card such as a LAN card and a modem, etc. The communication section 509 performs communication processing via a network such as the Internet. A drive 510 is also connected to I / O interface 505 as needed. A removable medium 511, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on drive 510 as needed so that computer programs read from it can be installed into storage section 508 as needed.

[0066] Specifically, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable storage medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 509, and / or installed from removable medium 511. When the computer program is executed by central processing unit (CPU) 501, it performs the functions defined in the methods of this application. It should be noted that the computer-readable storage medium of this application can be a computer-readable signal medium or a computer-readable storage medium or any combination thereof. The computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can also be any computer-readable storage medium other than a computer-readable storage medium that can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. Program code contained on a computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0067] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof. Programming languages ​​include object-oriented programming languages—such as Java, Smalltalk, and C++—as well as conventional procedural programming languages—such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0068] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0069] The modules described in the embodiments of this application can be implemented in software or in hardware.

[0070] In another aspect, this application also provides a computer-readable storage medium, which may be included in the electronic device described in the above embodiments; or it may exist independently and not assembled into the electronic device. The aforementioned computer-readable storage medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to: construct a multi-perspective sample generation engine, utilize a heterogeneous knowledge base and multi-source event corpus, and generate a set of multi-perspective answers for the same event sample through differentiated prompt templates; perform semantic deduplication and consistency verification on the multi-perspective answer set, calculate the semantic similarity between answers based on a pre-trained language model, perform clustering deduplication on answers with similarity exceeding a threshold, and verify the validity and timeliness of cited information through a knowledge graph; establish a difficulty quantification evaluation matrix, perform objective and subjective difficulty grading on training samples, and calculate a comprehensive difficulty score by combining the results of both; divide the training samples into multiple difficulty levels according to the comprehensive difficulty score, and perform progressive phased training based on the grading results, enabling the model to gradually transition from low-difficulty samples to high-difficulty samples, completing multi-perspective fusion learning and progressive capability optimization.

[0071] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A training method for an event analysis model based on progressive multi-perspective exploration, characterized in that, include: S1: Construct a multi-perspective sample generation engine, utilizing heterogeneous knowledge bases and multi-source event corpora, and generate a set of multi-perspective answers for the same event sample through differentiated prompt templates; S2: Perform semantic deduplication and consistency verification on the multi-perspective answer set, calculate the semantic similarity between answers based on the pre-trained language model, perform clustering deduplication on answers with similarity exceeding the threshold, and verify the validity and timeliness of the cited information through the knowledge graph; S3: Establish a difficulty quantification assessment matrix to classify the training samples using a combination of objective and subjective difficulty levels. The objective dimensions include semantic complexity, clause coverage, and inference chain complexity, while the subjective dimensions are determined based on model scores. The combined results of both are then used to calculate the overall difficulty score. S4: Based on the comprehensive difficulty score, the training samples are divided into multiple difficulty levels, and progressive phased training is performed based on the classification results, so that the model gradually transitions from low-difficulty samples to high-difficulty samples, completing multi-view fusion learning and progressive optimization of capabilities.

2. The event analysis model training method based on progressive multi-view exploration according to claim 1, characterized in that, The differentiated prompt templates include semantic parsing type, contextual analogy type, and element aggregation type templates.

3. The event analysis model training method based on progressive multi-view exploration according to claim 1, characterized in that, The multi-perspective sample generation engine calls multiple heterogeneous models, including a fine-tuned model for event analysis and a general large language model, and forms the multi-perspective answer set by comparing the semantic differences of the results generated by different models.

4. The event analysis model training method based on progressive multi-view exploration according to claim 1, characterized in that, The semantic deduplication process uses the BERT-legal model to calculate the semantic similarity between answers. Answers with similarity exceeding a threshold θ are clustered for deduplication, with the threshold θ set to 0.

85.

5. The event analysis model training method based on progressive multi-view exploration according to claim 1, characterized in that, The overall score for the objective dimensions is determined according to the following formula: D k = 0.4a + 0.3b + 0.3c, where a represents semantic complexity, b represents clause coverage, and c represents reasoning chain complexity.

6. The event analysis model training method based on progressive multi-view exploration according to claim 5, characterized in that, The overall difficulty score is composed of the objective score D. k Subjective score D z The overall difficulty score was determined to be D. final =0.6D k +0.4(D z / 100).

7. The event analysis model training method based on progressive multi-view exploration according to claim 1, characterized in that, The progressive phased training in S4 specifically includes: in the first stage, training is conducted based on low-difficulty samples to build the basic reasoning ability of the model; in the second stage, after the model's training on low-difficulty samples is stable, medium-difficulty samples are introduced for mixed training to enhance the ability; in the third stage, when the model's pass rate on medium-difficulty samples reaches a preset standard, the training ratio of high-difficulty samples is gradually increased based on the difficulty quantification evaluation results.

8. A computer-readable storage medium having one or more computer programs stored thereon, characterized in that, When the one or more computer programs are executed by a computer processor, they perform the method according to any one of claims 1-7.

9. A training system for an event analysis model based on progressive multi-perspective exploration, characterized in that, include: The multi-perspective sample generation unit is configured to build a multi-perspective sample generation engine. It utilizes a heterogeneous knowledge base and multi-source event corpus to generate a set of multi-perspective answers for the same event sample through differentiated prompt templates. The answer processing unit is configured to perform semantic deduplication and consistency verification on the multi-perspective answer set, calculate the semantic similarity between answers based on a pre-trained language model, perform clustering deduplication on answers with similarity exceeding a threshold, and verify the validity and timeliness of the reference information through a knowledge graph. The difficulty quantification assessment unit is configured to build a difficulty quantification assessment matrix, performing a combined objective and subjective difficulty classification on the training samples. The objective dimensions include semantic complexity, clause coverage, and inference chain complexity, while the subjective dimensions are determined based on model scores. The overall difficulty score is calculated by combining the results of both approaches. The progressive training unit is configured to divide the training samples into multiple difficulty levels based on the comprehensive difficulty score, and perform progressive phased training based on the classification results, so that the model gradually transitions from low-difficulty samples to high-difficulty samples, completing multi-view fusion learning and progressive optimization of capabilities.

10. The event analysis model training system based on progressive multi-view exploration according to claim 9, characterized in that, The differentiated prompt templates include semantic parsing type, contextual analogy type, and element aggregation type templates; the multi-perspective sample generation engine calls multiple heterogeneous models, including event analysis domain fine-tuning models and general large language models, and forms the multi-perspective answer set by comparing the semantic differences of the results generated by different models.

11. The event analysis model training system based on progressive multi-view exploration according to claim 9, characterized in that, The semantic deduplication process uses the BERT-legal model to calculate the semantic similarity between answers. Answers with similarity exceeding a threshold θ are clustered for deduplication, with the threshold θ set to 0.

85.

12. The event analysis model training system based on progressive multi-view exploration according to claim 9, characterized in that, The overall difficulty score is composed of the objective score D. k Subjective score D z The overall difficulty score was determined to be D. final =0.6D k +0.4(D z / 100), the comprehensive score of the objective dimension is determined according to the following formula: D k = 0.4a + 0.3b + 0.3c, where a represents semantic complexity, b represents clause coverage, and c represents reasoning chain complexity.

13. The event analysis model training system based on progressive multi-view exploration according to claim 9, characterized in that, The progressive phased training specifically includes: in the first phase, training is conducted based on low-difficulty samples to build the model's basic reasoning ability; in the second phase, after the model's training on low-difficulty samples has stabilized, medium-difficulty samples are introduced for mixed training to enhance the ability; in the third phase, when the model's pass rate on medium-difficulty samples reaches a preset standard, the training ratio of high-difficulty samples is gradually increased based on the difficulty quantification evaluation results.

Images