Long text understanding ability testing method based on context association

Abstract

The invention relates to the technical field of natural language processing, in particular to a long text understanding ability testing method and platform based on context association. The method comprises the steps of obtaining a long text data set and labeling a logic relation, constructing a context correlation analysis model to extract entities and events, designing a dynamic test case to simulate a complex scene, inputting a model to be tested to obtain a comprehension result, quantitatively evaluating model performance and the like. According to the method, test cases are optimized through multi-modal information integration and adversarial training, the context understanding ability of the model in a complex environment is comprehensively evaluated, and a real-time feedback mechanism is provided for guiding optimization. According to the method, the problems of single test case, insufficient quantitative evaluation and insufficient multi-modal integration in the prior art can be effectively solved, and a scientific basis is provided for iterative optimization of a large model.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of natural language processing, and specifically relates to a long text understanding ability test method and platform based on context association. BACKGROUND

[0002] In the field of long text understanding ability testing, by analyzing the context association and semantic integration ability of the to-be-tested large model, the cognitive performance of complex text processing can be significantly improved. Among them, the context understanding ability of the model usually depends on its accurate capture of cross-paragraph logical relationships and semantic associations. In traditional methods, the annotation of text data sets and the design of test cases are often limited to a single scene or fixed mode, which makes it difficult for test results to fully reflect the performance of the model in real complex environments. Therefore, the existing technology needs a method that can simulate diversified scenarios and quantitatively evaluate the long text understanding ability of the model.

[0003] Long text understanding ability testing based on context association is an emerging technical means. This method can effectively simulate the complexity in real scenarios by constructing dynamic test cases, such as inserting interference information, deleting key paragraphs, or modifying semantic associations. At the same time, by combining knowledge graph technology and graph neural networks, entities, events, and logical relationships in long text can be mapped into graph structures, enabling dynamic updating and precise analysis of context-dependent relationships. This method not only improves the diversity and challenge of test cases, but also evaluates the cross-modal understanding ability of the model in multi-modal scenarios, thereby better adapting to actual application needs.

[0004] The existing long text understanding ability testing method has the following limitations: Single test cases: Current test case design is mostly based on fixed patterns, lacking sufficient consideration of noise data and logical relationship changes, resulting in test results that may not fully reflect the actual performance of the model.

[0005] Deficiency of quantitative evaluation indicators: Existing evaluation methods usually focus on single-dimensional performance indicators, while comprehensive consideration of multi-dimensional indicators such as context consistency, semantic integrity, and anti-interference performance is limited.

[0006] Lack of multi-modal integration: In cross-modal scenarios, existing methods rarely involve joint analysis of long text and image, audio, or video data, which limits the verification of model applicability in complex multi-modal tasks.

[0007] This invention aims to improve the comprehensiveness and accuracy of long text comprehension testing by introducing dynamic test case design and multimodal information integration testing, combined with adversarial training to optimize test cases. This method can more effectively evaluate the model's contextual understanding ability in complex environments and provide a reliable basis for the iterative optimization of large models. Summary of the Invention

[0008] The purpose of this invention is to provide a method and platform for testing long-text comprehension capabilities based on contextual association. By constructing dynamic test cases, quantifying evaluation metrics, and integrating multimodal information, this method solves the problems of test case uniformity, insufficient quantitative evaluation, and lack of multimodal integration in existing technologies. This method can comprehensively reflect the model's contextual understanding capability in complex environments and provide a reliable basis for the iterative optimization of large models.

[0009] To achieve the above objectives, this invention provides a method for testing long text comprehension ability based on contextual association, comprising the following steps: Step S1: Obtain a long text dataset containing multiple paragraphs and label its contextual logical relationships and semantic associations; Step S2: Construct a context association analysis model to extract entities, events, and cross-paragraph dependencies from the text; Step S3: Design dynamic test cases, including inserting interfering information, deleting key paragraphs, or modifying semantic relationships, to simulate the complexity of real-world scenarios; Step S4: Input the test cases into the large model to be tested and obtain the understanding results output by the model; Step S5: Quantitatively evaluate the model's ability to understand long texts by comparing the contextual relationship between the model output and the annotation.

[0010] Preferably, step S1 specifically includes: Acquire a long-text dataset covering various fields, including at least scientific literature, long news reports, and fictional narrative texts. Each long text is segmented and annotated to clarify the logical relationships and semantic connections between paragraphs. For example, in scientific literature, annotate logical structures such as causal relationships, temporal relationships, and conditional relationships; in long news reports, annotate the timeline of events and the evolution of themes; and in fictional narrative texts, annotate character relationships and plot progression clues. The annotation adopts a hierarchical structure: the first layer defines the overall logical framework between paragraphs; the second layer refines the semantic connections between sentences within a paragraph; and the third layer annotates the contextual meaning changes of specific words.

[0011] Preferably, step S2 specifically includes: A contextual association analysis model is constructed, employing knowledge graph technology to map entities, events, and logical relationships in long texts into a graph structure. Specifically, firstly, entities and events are extracted from the text as nodes, and noun phrases, verb phrases, and their modifiers are identified using natural language processing tools. Then, edges are established between nodes based on the labeled logical and semantic relationships. For example, causal relationships are represented by directed edges, temporal relationships by edges with timestamps, and conditional relationships by weighted edges. Finally, the graph structure is dynamically updated using a graph neural network. Each update operation includes information aggregation of adjacent nodes and feature propagation across the global graph, thereby capturing cross-paragraph dependencies.

[0012] Preferably, step S3 specifically includes the following steps: Designing dynamic test cases to simulate the complexity of real-world scenarios can be done in three ways: (1) Randomly insert irrelevant paragraphs into long texts, with the insertion position chosen based on the strength of the logical relationship between paragraphs. For example, insert a paragraph unrelated to the topic between paragraphs with strong causal relationships and observe whether the model can ignore the interfering information and correctly understand the logical chain.

[0013] (2) Delete core paragraphs and select paragraphs that are crucial to understanding the context to remove, such as removing intermediate links in the causal chain or key turning points in the development of events, and evaluate the model’s reasoning ability for missing information.

[0014] (3) Modify the causal or temporal relationships between paragraphs, such as adjusting the time sequence of events or changing the conditions for the conditional relationship to verify the model’s sensitivity to logical associations.

[0015] Preferably, step S4 specifically includes: Input the designed dynamic test cases into the large model under test and record the model's understanding results. The model output includes a summary of the theme of each paragraph, inferences about the logical relationships between paragraphs, and a summary of the core semantics of the entire text. For multimodal test cases, the model also needs to output an interpretation of the association between image, audio, or video data. For example, in a long text containing news images, the model needs to supplement its understanding of the text description by combining the image content.

[0016] Preferably, step S5 specifically includes: The model's ability to understand long texts is quantitatively evaluated by comparing the contextual relevance of the model's output with the labeled text. The quantitative evaluation includes the following metrics: (1) Context consistency score, calculated using the following formula: Where Ncorrect represents the number of logical relationships in the model output that are consistent with the annotation, and Ntotal represents the total number of logical relationships.

[0017] (2) Semantic integrity score, calculated using the following formula:

[0018] Where oi is the semantic summary of the i-th segment output by the model, gi is the semantic summary of the i-th segment labeled, sim(⋅) is the semantic similarity function, and wi is the segment weight, which is determined by the importance of the segment in the whole text.

[0019] (4) Anti-interference performance score, calculated using the following formula:

[0020] Where Nresist represents the number of times the model successfully resists interference information, and Ndisturb represents the total number of interference information.

[0021] Furthermore, the method also includes multimodal information integration testing, specifically: Multimodal test cases are constructed by associating long texts with image, audio, or video data. For example, photos of experimental equipment can be added to scientific literature, on-site audio clips to long news reports, and video clips of character dialogues to fictional narratives. The model needs to combine multimodal information to generate a comprehensive understanding of the text. Evaluation metrics include cross-modal consistency score and multimodal semantic integrity score.

[0022] Furthermore, the method also includes adversarial training to optimize test cases, specifically: Targeted adversarial examples are generated based on the weaknesses in the model's output. For example, if the model performs poorly in handling causal relationships, more causal-related test cases are added; if the model is not sensitive enough to temporal relationships, more perturbation tests involving temporal relationships are designed. The difficulty and coverage of the test cases are iteratively adjusted until the model performance reaches a preset threshold.

[0023] Furthermore, the method also includes a real-time feedback mechanism, specifically: Visualize the test results, highlighting contextual fragments where the model misunderstands. For example, use colors to mark incorrect logical relationships and arrows to indicate missing inference chains. Simultaneously, provide corrective suggestions to guide model parameter adjustments or structural optimization. For instance, for models that misunderstand causal relationships, suggest increasing training data on causal relationships or adjusting the weight allocation of the attention mechanism.

[0024] By adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art. Of course, any product implementing the present invention does not necessarily need to achieve all of the following advantages at the same time: This invention employs the aforementioned context-based long text understanding capability testing method. Through the design of dynamic test cases and the application of multi-dimensional quantitative evaluation metrics, it significantly improves the understanding of the model's performance in processing complex text. Furthermore, by combining multimodal information integration testing and adversarial training to optimize test cases, it can more comprehensively evaluate the model's context understanding capability in diverse scenarios, providing a scientific basis and technical support for the iterative optimization of large models.

[0025] The specific embodiments of the present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0026] The accompanying drawings described below are merely some embodiments. Those skilled in the art can obtain other drawings based on these drawings without any creative effort. In the drawings: Figure 1 This is a flowchart of the long text comprehension ability testing method based on context association of the present invention.

[0027] It should be noted that these accompanying drawings and textual descriptions are not intended to limit the scope of the invention in any way, but rather to illustrate the concept of the invention to those skilled in the art by referring to specific embodiments. Detailed Implementation

[0028] The invention will now be described in further detail with reference to the accompanying drawings.

[0029] Please see Figure 1 As shown, this invention provides a method and platform for testing long text comprehension ability based on context association, including the following steps: First, in step S1, acquiring the long text dataset is the fundamental operation of the entire testing method. The sources of the long text dataset can include multiple fields such as scientific literature, long news reports, and fictional narrative texts to ensure data diversity and representativeness. In the specific implementation process, each long text needs to be segmented and annotated to clarify the logical relationships and semantic connections between paragraphs. For example, in scientific literature, logical structures such as causal relationships, temporal relationships, and conditional relationships are annotated; in long news reports, the timeline of event development and thematic evolution are annotated; and in fictional narrative texts, character relationships and plot progression clues are annotated. The annotation adopts a hierarchical structure: the first layer defines the overall logical framework between paragraphs, the second layer refines the semantic connection points between sentences within a paragraph, and the third layer annotates the contextual meaning changes of specific words. This annotation information, generated manually or automatically, forms an annotated database containing multiple long text paragraphs and their contextual logical relationships and semantic connections. This database serves as the basic input for subsequent testing, and its accuracy and completeness directly affect the reliability of the test results.

[0030] Next, in step S2, a context association analysis model is constructed to extract entities, events, and cross-paragraph dependencies from the text. The core of this model lies in the application of knowledge graph technology. Specifically, entities and events are first extracted as nodes from the long text dataset, and noun phrases, verb phrases, and their modifiers are identified using natural language processing tools. Then, edges are established between nodes based on the labeled logical and semantic relationships. For example, causal relationships are represented by directed edges, temporal relationships by timestamped edges, and conditional relationships by weighted edges. Throughout this process, the construction of nodes and edges must strictly adhere to the logical framework and semantic connection points in the labeled database to ensure the accuracy of the graph structure. Finally, the graph structure is dynamically updated using a graph neural network. Each update operation includes information aggregation of adjacent nodes and feature propagation of the global graph, thereby capturing cross-paragraph dependencies. The output of this model is a knowledge graph that reflects the context associations of long texts, providing an important basis for the design of subsequent dynamic test cases.

[0031] In step S3, dynamic test cases are designed to simulate the complexity of real-world scenarios. The design of dynamic test cases is based on the output of the context association analysis model and includes three methods. The first method involves randomly inserting irrelevant paragraphs into long texts, with the insertion position chosen based on the strength of the logical relationship between paragraphs. For example, inserting a piece of content unrelated to the topic between paragraphs with strong causal relationships allows observation of whether the model can ignore interfering information and correctly understand the logical chain. The second method involves deleting core paragraphs, selecting paragraphs crucial for contextual understanding for removal, such as removing intermediate links in causal chains or key turning points in event development, to evaluate the model's reasoning ability regarding missing information. The third method involves modifying the causal or temporal relationships between paragraphs, such as adjusting the chronological order of events or changing the conditions for conditional relationships, to verify the model's sensitivity to logical associations. The design of these dynamic test cases needs to incorporate node and edge information from the knowledge graph to ensure the complexity and diversity of the test cases.

[0032] In step S4, the designed dynamic test cases are input into the large model under test, and the model's understanding results are recorded. The input to the large model under test includes the long text in the dynamic test cases and its related annotation information. The model output includes a summary of the theme of each paragraph, an inference of the logical relationships between paragraphs, and a summary of the core semantics of the entire text. For multimodal test cases, the model also needs to output an interpretation of the association between image, audio, or video data. For example, in a long text containing news images, the model needs to supplement its understanding of the text description by combining the image content. During this process, the model's output results need to be compared with the logical framework and semantic connection points in the annotation database to evaluate its ability to capture contextual relationships.

[0033] In step S5, the model's long text understanding capability is quantitatively evaluated by comparing the contextual relevance between the model output and the annotations using the quantitative evaluation module. The evaluation metrics for the quantitative evaluation module include contextual consistency score, semantic integrity score, and robustness score. The formula for calculating the contextual consistency score is: Where Ncorrect represents the number of logical relations in the model output that are consistent with the annotation, and Ntotal represents the total number of logical relations. The formula for calculating the semantic integrity score is: Where oi is the semantic summary of the i-th segment output by the model, gi is the semantic summary of the i-th segment annotated, sim(⋅) is the semantic similarity function, and wi is the segment weight, determined by the importance of the segment in the whole text. The formula for calculating the anti-interference performance score is: Here, Nresist represents the amount of interference information the model successfully resists, and Ndisturb represents the total amount of interference information. These quantitative evaluation metrics comprehensively reflect the model's contextual understanding ability in complex environments through specific numerical calculations.

[0034] Furthermore, this invention also includes multimodal information integration testing. Specifically, long texts are associated with image, audio, or video data to construct multimodal test cases. For example, photos of experimental equipment are added to scientific literature, on-site audio clips are added to long news reports, and video clips of character dialogues are added to fictional narrative texts. The model needs to combine multimodal information to generate a comprehensive understanding of the text. Evaluation metrics include cross-modal consistency score and multimodal semantic integrity score. The cross-modal consistency score measures the model's ability to capture the associations between different modalities of information, while the multimodal semantic integrity score evaluates the model's comprehensive understanding of multimodal information. This testing process needs to incorporate the output of the quantitative evaluation module to ensure the scientific validity and objectivity of the evaluation metrics.

[0035] Furthermore, this invention also includes adversarial training to optimize test cases. Targeted adversarial examples are generated based on weaknesses in the model's output. For example, if the model performs poorly in handling causal relationships, causal-related test cases are added; if the model lacks sensitivity to temporal relationships, more perturbation tests involving temporal relationships are designed. The difficulty and coverage of the test cases are iteratively adjusted until the model performance reaches a preset threshold. This process requires adherence to dynamic test case design principles to ensure the diversity and challenge of the adversarial examples.

[0036] Finally, this invention also includes a real-time feedback mechanism. The test results are visualized, highlighting contextual fragments where the model misunderstands. For example, incorrect logical relationships are marked with colors, and missing inference chains are indicated with arrows. Simultaneously, corrective suggestions are provided to guide model parameter adjustments or structural optimization. For instance, for models that misunderstand causal relationships, suggestions may include increasing training data on causal relationships or adjusting the weight allocation of the attention mechanism. This mechanism is implemented through the output of the quantification evaluation module, ensuring the accuracy and operability of the feedback information.

[0037] The above embodiments describe in detail the specific implementation process of the present invention. Through the synergistic cooperation of the above steps, the present invention achieves a comprehensive test of long text comprehension capabilities, providing a reliable basis for the iterative optimization of large models.

[0038] To enable those skilled in the art to fully understand and implement this invention, the specific implementation principle of the invention is further explained below with reference to a specific application scenario. This scenario uses a multimodal test of long news reports as an example to elaborate in detail the actual operation process and technical implementation principle of a long text comprehension ability testing method based on contextual association.

[0039] In step S1, long-form reports containing multiple events are first obtained from news organizations as a long-text dataset. These reports cover multiple fields such as international relations, natural disasters, and technological breakthroughs, ensuring the diversity and representativeness of the data. For example, in a report about an international summit, the annotation tool extracts the logical relationships between paragraphs, clarifying semantic connections such as time sequence, causal relationships, and conditional relationships. Specifically, the annotation tool defines "a leader of a certain country makes a statement" and "another country subsequently responds" as a causal relationship, and marks "the start time of the meeting" and "the signing time of the agreement" as a temporal relationship. This annotation information is stored in a hierarchical structure. The first layer records the overall logical framework between paragraphs, such as the timeline of event development; the second layer refines the semantic connection points between sentences within a paragraph, such as the behavior and statements of specific individuals; the third layer annotates the contextual meaning changes of words, such as the specific reference of the word "statement" in different contexts. The resulting annotation database provides a reliable basic input for the design of subsequent dynamic test cases.

[0040] In step S2, a contextual association analysis model is constructed to extract entities, events, and cross-paragraph dependencies from the text. In this scenario, the model employs knowledge graph technology to map entities (such as "a country's leader," "international summit") and events (such as "issuing a statement," "signing an agreement") in the report as nodes. After natural language processing tools identify noun phrases and verb phrases, edges between nodes are established based on logical relationships in the labeled database. For example, a directed edge represents a causal relationship between "a country's leader issuing a statement" and "another country subsequently responding," while a timestamped edge represents a temporal relationship between "the start time of the meeting" and "the signing time of the agreement." Subsequently, a graph neural network dynamically updates the graph structure, with each update operation including information aggregation of adjacent nodes and feature propagation of the global graph. This process enables the model to capture cross-paragraph dependencies, such as how "a country's leader's statement" affects the subsequent "signing of an international agreement."

[0041] In step S3, dynamic test cases are designed to simulate the complexity of real-world scenarios. For the aforementioned news reports, firstly, irrelevant content is inserted between paragraphs with strong causal relationships. For example, a description of the local weather is inserted between "a leader of a certain country issued a statement" and "another country subsequently responded," to observe whether the model can ignore interfering information and correctly reason the logical chain. Secondly, key paragraphs are deleted, such as removing content related to "a leader of a certain country issued a statement," to evaluate the model's reasoning ability regarding missing information. Finally, the logical relationships between paragraphs are modified, such as adjusting the chronological order of events or changing causal relationships to conditional relationships, to verify the model's sensitivity to logical associations. These dynamic test cases are all designed based on node and edge information in the knowledge graph, ensuring the complexity and diversity of the test cases.

[0042] In step S4, the designed dynamic test cases are input into the large model under test, and the model's output understanding results are recorded. For example, the model needs to summarize the theme of each paragraph, infer the logical relationships between paragraphs, and summarize the core semantics of the entire text. For multimodal test cases, the model also needs to combine relevant image or audio data to generate a comprehensive understanding result. In the above scenario, the model needs to combine photos and on-site audio clips of the international summit to supplement its understanding of the text description. For example, the model needs to infer whether the atmosphere of the meeting is tense or friendly based on the venue layout and the expressions of the participants in the photos, and further refine its interpretation of the event development by combining the tone and intonation in the audio clips. In this process, the model's output results are compared with the logical framework and semantic connection points in the labeled database to evaluate its ability to capture contextual relationships.

[0043] In step S5, the model's long text understanding ability is quantitatively evaluated by comparing the contextual relevance between the model output and the annotations using a quantitative evaluation module. Specifically, the contextual consistency score is calculated as the proportion of logical relationships consistent between the model output and the annotations to the total number of logical relationships. For example, in the above scenario, if the model correctly identifies the causal relationship between "a leader of a certain country makes a statement" and "another country subsequently responds," it is counted as a correct logical relationship. The semantic integrity score is calculated by using a semantic similarity function to determine the similarity between the paragraph summary output by the model and the annotation summary, and then combining this with paragraph weights to arrive at a comprehensive score. The anti-interference performance score is calculated as the proportion of the number of times the model successfully resists interference information to the total number of interference information. For example, if the model can still correctly infer causal relationships even when weather descriptions are inserted, it is counted as a successful anti-interference case.

[0044] Furthermore, for multimodal information integration testing, news reports are associated with relevant image and audio data to construct multimodal test cases. For example, photos and audio clips of international summits are added to the reports, requiring the model to combine multimodal information to generate a comprehensive understanding. Evaluation metrics include cross-modal consistency score and multimodal semantic integrity score. The cross-modal consistency score measures the model's ability to capture the correlation between different modalities of information; for example, whether the model can infer the importance and atmosphere of the meeting based on the venue layout in the photos. The multimodal semantic integrity score evaluates the model's comprehensive understanding of multimodal information; for example, whether the model can combine the tone and intonation in the audio clips to improve the interpretation of the event's development.

[0045] Simultaneously, targeted adversarial examples are generated based on the weaknesses in the model's output. For example, if the model performs poorly in handling causal relationships, more causal-related test cases are added; if the model is not sensitive enough to temporal relationships, more perturbation tests involving temporal relationships are designed. The difficulty and coverage of the test cases are iteratively adjusted until the model performance reaches a preset threshold. This process requires adherence to dynamic test case design principles to ensure the diversity and challenge of the adversarial examples.

[0046] Finally, the test results are visualized through a real-time feedback mechanism. For example, incorrect logical relationships are marked with colors, and missing inference chains are indicated with arrows. Simultaneously, corrective suggestions are provided to guide model parameter adjustments or structural optimization. For instance, for models that misunderstand causal relationships, suggestions include increasing training data on causal relationships or adjusting the weight allocation of the attention mechanism. This mechanism is implemented through the output of the quantitative evaluation module, ensuring the accuracy and operability of the feedback information.

[0047] The implementation principle of this invention has been further explained above in conjunction with specific application scenarios. Through the synergistic cooperation of the above steps, this invention achieves comprehensive testing of complex texts such as long news reports, significantly improves the understanding of the model's contextual understanding ability, and provides a scientific basis and technical support for the iterative optimization of large models.

[0048] This invention is not limited to the embodiments described above. Anyone should understand that structural changes made under the guidance of this invention, and any technical solutions that are the same as or similar to this invention, fall within the protection scope of this invention. Technical aspects, shapes, and structures not described in detail in this invention are all publicly known technologies.

Claims

1. A method and platform for testing long text comprehension ability based on context association, characterized in that, Includes the following steps: Step S1: Obtain a long text dataset containing multiple paragraphs and label the logical relationships and semantic connections between each paragraph; Step S2: Construct a context association analysis model to extract entities, events, and cross-paragraph dependencies from the text; Step S3: Design dynamic test cases, including inserting distracting information, deleting key paragraphs, or modifying semantic relationships; Step S4: Input the test cases into the large model to be tested and obtain the understanding results output by the model; Step S5: Quantitatively evaluate the model's ability to understand long texts by comparing the contextual relationship between the model output and the annotation.

2. The method and platform for testing long text comprehension ability based on context association as described in claim 1, characterized in that, Step S1 specifically includes: Acquire a long text dataset covering scientific literature, long news reports, and fictional narrative texts; segment and annotate each long text to clarify the logical relationships and semantic connections between paragraphs; the annotation adopts a hierarchical structure, with the first layer defining the overall logical framework between paragraphs, the second layer refining the semantic connection points between sentences within a paragraph, and the third layer annotating the contextual meaning changes of specific words.

3. The method and platform for testing long text comprehension ability based on context association as described in claim 1, characterized in that, Step S2 specifically includes: A context association analysis model is constructed, which uses knowledge graph technology to map entities, events, and logical relationships in long texts into a graph structure. First, entities and events are extracted from the text as nodes, and noun phrases, verb phrases, and their modifiers are identified using natural language processing tools. Then, edges between nodes are established based on the labeled logical relationships and semantic associations. Finally, the graph structure is dynamically updated using a graph neural network, with each update operation including the aggregation of information from adjacent nodes and the propagation of features from the global graph.

4. The method and platform for testing long text comprehension ability based on context association according to claim 1, characterized in that, Step S3 specifically includes: Randomly insert irrelevant paragraphs into long texts, with the insertion position chosen based on the strength of the logical relationship between paragraphs; delete core paragraphs, selecting those crucial to understanding the context for removal; modify causal or temporal relationships between paragraphs, adjusting the chronological order of events or changing the conditions for conditional relationships.

5. The method and platform for testing long text comprehension ability based on context association according to claim 1, characterized in that, Step S4 specifically includes: Input the designed dynamic test cases into the large model to be tested and record the understanding results output by the model. The model output includes a summary of the theme of each paragraph, an inference of the logical relationship between paragraphs, and a summary of the core semantics of the whole text. For multimodal test cases, the model also needs to output the related interpretation of image, audio or video data.

6. The method and platform for testing long text comprehension ability based on context association according to claim 1, characterized in that, Step S5 specifically includes: The context consistency score, semantic integrity score, and robustness score are calculated using the quantitative evaluation module; the formula for calculating the context consistency score is as follows: The formula for calculating semantic integrity score is: The formula for calculating the anti-interference performance score is: .

7. The method and platform for testing long text comprehension ability based on context association according to claim 1, characterized in that, It also includes multimodal information integration testing, specifically: Associate long texts with image, audio, or video data to construct multimodal test cases; the model needs to combine multimodal information to generate a comprehensive understanding of the text; evaluation metrics include cross-modal consistency score and multimodal semantic integrity score.

8. The method and platform for testing long text comprehension ability based on context association according to claim 1, characterized in that, This also includes adversarial training to optimize test cases, specifically: Based on the weaknesses in the model output, targeted adversarial examples are generated; if the model performs poorly in handling causal relationships, test cases related to causal relationships are added; if the model is not sensitive enough to temporal relationships, more perturbation tests of temporal relationships are designed; the difficulty and coverage of test cases are adjusted iteratively until the model performance reaches the preset threshold.

9. The method and platform for testing long text comprehension ability based on context association according to claim 1, characterized in that, It also includes a real-time feedback mechanism, specifically: Visualize the test results, highlighting contextual fragments where the model misunderstands; use color to mark erroneous logical relationships and arrows to indicate missing inference chains; provide corrective suggestions to guide model parameter adjustments or structural optimization.

10. The method and platform for testing long text comprehension ability based on context association according to claim 1, characterized in that, The long text dataset includes scientific and technological literature, long news reports, and fictional narrative texts, and the labeled content includes logical structures such as causal relationships, temporal relationships, and conditional relationships.

Images