Method and system for evaluating information pollution in large model

Abstract

The invention provides an information pollution assessment method and system in a large model, and the method comprises the steps: obtaining a pollution corpus for the training of the large model, and constructing an input pollution vector for each corpus in the pollution corpus; generating an output set by using the large model, and constructing an output pollution vector for each piece of output content in the output set; according to the actual semantics of the multiple corpora and the multiple output contents, the corpora and the output contents with the actual semantics meeting the similarity requirement are matched, and the inheritance pollution amount between each set of output contents and the corpora is calculated; calculating a pollution influence coefficient according to the output pollution vector and the inherited pollution amount; constructing nodes in an application system of the large model, and recording node risk values of output contents; and calculating a global pollution risk index according to the input pollution vector, the output pollution vector and the node risk value. The problem that negative results or wrong results are generated after a large model is influenced by polluted content is solved.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to methods and systems for assessing information pollution in large models. Background Technology

[0002] In recent years, with the large-scale application of large-scale models in scenarios such as search, question answering, content generation, code assistance, and decision support, the dependence of these models on training data and operating environments has continued to deepen. Large-scale models typically learn language patterns, knowledge structures, and reasoning models automatically from massive internet corpora. The authenticity, logic, and security of their generated content largely depend on the quality and distribution structure of the training corpus. However, open internet data sources are complex, and the corpus contains unverified false reports, exaggerated narratives, emotional texts, outdated information, stereotypical and biased expressions, copyright-restricted content, and potentially illegal data. Once these low-quality or biased corpora are absorbed by the model, they will form biased representations or erroneous associations in the parameter space, which will then manifest in various forms during the reasoning and generation stages, such as factual errors, logical inconsistencies, semantic ambiguity, emotional misleading, and value bias, weakening the reliability and compliance of the model's output. On the other hand, the output of large language models is not a "disposable product," but is reused and amplified multiple times in actual platforms. It is repeatedly used and spread in the closed-loop chain of "training corpus → model generation → system storage and indexing → retraining / re-exposure," causing pollution to accumulate in the system and be difficult to eliminate.

[0003] Therefore, there is an urgent need for a method to assess the pollution risk in large models and reduce the negative impact of pollution information on the reasoning logic and generated content of large models. Summary of the Invention

[0004] In view of this, the purpose of this application is to propose a method and system for assessing information pollution in large models, which solves the problem of generating negative or erroneous results after large models are affected by polluted content.

[0005] To achieve one of the aforementioned objectives, this application provides a method for assessing information pollution in a large-scale model, the method comprising: Obtain a polluted corpus for training the large model, and construct an input pollutant vector for each piece of text in the polluted corpus; The polluted corpus is input into the large model, and the large model is used to generate an output set. An output pollution vector is constructed for each output content in the output set. Based on the actual semantics of multiple corpora and multiple output contents, match the corpora and output contents whose actual semantics meet the similarity requirements, and calculate the inheritance pollution amount between each set of matched output contents and the corpora; Based on the output pollution vector and the inherited pollution amount, calculate the pollution impact coefficient of the large model; Nodes are constructed in the application system of the large model, and the nodes are used to record the node risk value of the output content in the application system. Based on the multiple input contamination vectors, the multiple output contamination vectors, and the multiple node risk values, a global contamination risk index is calculated. The contamination impact coefficient and the global contamination risk index characterize the degree of information contamination in the large model.

[0006] As a further improvement to one embodiment of this application, the step of matching the corpus and the output content whose actual semantics meet the similarity requirements based on the actual semantics of the plurality of corpora and the plurality of output contents includes: Using the same semantic encoding, multiple corpora and multiple output contents are mapped to the same semantic space; Calculate the similarity between the multiple corpora and the multiple output contents; For each output content, match the corpus with the highest similarity.

[0007] As a further improvement to one embodiment of this application, the calculation of the amount of inherited pollution between the output content after each matching group and the corpus includes: The weight of each output content matching the corpus is calculated according to the following formula: ; Where, α j，i The weights of the corpus i that match the output content j, where λ represents the degree of concentration of the weight distribution, and h i The corpus after semantic encoding The output content after semantic encoding, d i ∈N(y j The corpus that meets the similarity requirements; The amount of inherited contamination is calculated using the following formula: ; in, The amount of inherited contamination. The input contamination vector; The step of calculating the pollution impact coefficient of the large model based on the output pollution vector and the inherited pollution amount includes: The pollution impact coefficient is calculated using the following formula: ; Among them, A j The pollution impact coefficient is... Let ε be the output contamination vector, and let ε be a constant to prevent the denominator from being zero.

[0008] As a further improvement to one embodiment of this application, the step of constructing nodes in the application system of the large model and using the nodes to record the node risk value of the output content in the application system includes: The nodes are used to record the dynamic behavior information of the output content in the system, including the first write time, the most recent access time, and the number of calls. Calculate the node risk value using the following formula: ; ; Specifically, The node risk value is the value calculated for the dimension in question. Let m be the node, and m be the node ordinal number. The most recent access time, The first write time is k, where k is the dimension. The pollution vector is in dimension 1. For node lifespan, This represents the number of times the function is called.

[0009] As a further improvement to one embodiment of this application, the step of calculating the global contamination risk index based on a plurality of input contamination vectors, a plurality of output contamination vectors, and a plurality of node risk values ​​includes: Calculate the mean input contamination of multiple input contamination vectors of multiple corpora, calculate the mean output contamination of multiple output contamination vectors of multiple output contents, and calculate the mean system node risk of multiple node values ​​of multiple output contents; The comprehensive risk index is calculated using the following formula: ; Among them, R k This refers to the comprehensive risk index. The input pollution mean, The average value of the output pollution. β1, β2, and β3 are the average values ​​of the system nodes, β3 are the scene coefficients, and k is the dimension. The global pollution risk index is calculated using the following formula: ; Where R global Let n be the global pollution risk index, and w be the total number of dimensions. k The weights are for dimension k.

[0010] As a further improvement to one embodiment of this application, the step of constructing an input pollution vector for each piece of text in the pollution corpus includes: For each piece of text in the polluted corpus, construct a multi-dimensional input pollution vector; The step of constructing an output contamination vector for each output item in the output set includes: For each output item in the output set, construct a multi-dimensional output contamination vector.

[0011] As a further improvement of one embodiment of this application, the input contamination vector and the output contamination vector of multiple dimensions all include a fact distortion risk dimension, a logical structure risk dimension, a semantic quality risk dimension, and a sentiment bias risk dimension. The input contamination vector or the output contamination vector in the fact distortion risk dimension is calculated according to the following formula: ; in, For the input contamination vector or the output contamination vector of the fact distortion risk dimension, d i For the corpus, Let h, r, and t be the set of triples in the corpus, where h, r, and t are the head entity, relation, and tail entity in the triple, respectively. The consistency score for triples; The corpus is split into sentence sequences, and the input contamination vector or the output contamination vector in the logical structure risk dimension is calculated according to the following formula: ; ; in, The input contamination vector or the output contamination vector is the risk dimension of the logical result. For the sentence sequence, The probability of a contradiction between two adjacent sentences; The input contamination vector or the output contamination vector in the semantic quality risk dimension is calculated according to the following formula: ; in, For the input contamination vector or the output contamination vector of the semantic quality risk dimension, q syn The syntactic structure score of the corpus is given by q. coref The reference consistency score is assigned to the corpus. The input contamination vector or the output contamination vector in the emotional bias wind dimension is calculated according to the following formula: ; in, For the input contamination vector or the output contamination vector of the emotional bias wind dimension, e(d) i b(d) represents the intensity of emotion. i () represents the intensity of biased expression.

[0012] As a further improvement to one embodiment of this application, after characterizing the information pollution level of the large model by the pollution impact coefficient and the global pollution risk index, the following is included: Set a pollution risk index threshold; when the global pollution risk index is greater than the pollution risk index threshold, the output content will be blocked. Analyze the knowledge reference links of the intercepted output content, query the input content corresponding to the output content, and correct the input content.

[0013] As a further improvement to one embodiment of this application, after characterizing the information pollution level of the large model by the pollution impact coefficient and the global pollution risk index, the following is included: In response to the numerical characterization of the pollution impact coefficient indicating that the large model amplifies the pollution content of the corpus, the inference logic of the large model is corrected.

[0014] Based on the same inventive concept, this application also provides an information pollution assessment system in a large model, the system comprising: The input risk module is used to obtain a pollution corpus for training the large model and to construct an input pollution vector for each piece of text in the pollution corpus. The output risk module is used to input the pollution corpus into the large model, use the large model to generate an output set, and construct an output pollution vector for each output content in the output set. The first matching module is used to match the actual semantics of the multiple corpora and the multiple output contents with the corpora whose actual semantics meet the similarity requirements, and to calculate the amount of inheritance pollution between the output contents and the corpora after each matching; The first calculation module is used to calculate the pollution impact coefficient of the large model based on the output pollution vector and the inherited pollution amount; The system risk module constructs nodes in the application system of the large model and uses these nodes to record the node risk values ​​of the output content in the application system. The evaluation module is used to calculate a global pollution risk index based on multiple input pollution vectors, multiple output pollution vectors, and multiple node risk values. The pollution impact coefficient and the global pollution risk index characterize the degree of information pollution in the large model.

[0015] Compared to existing technologies, the technical advantages of this invention lie in the following: by simultaneously vectorizing the pollutant materials used to train the large model and the output content generated during the training of the large model, and then combining this with the application links within the system, a complete evolution path of pollution from production to dissemination to reuse is constructed. This forms a multi-source fusion pollution risk index based on risks at the training end, generation end, and internal circulation risks within the system application, thereby achieving a closed-loop governance system covering the entire lifecycle of data, the large model, and the application system. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in this application or related technologies, the drawings used in the description of the implementation methods or related technologies will be briefly introduced below. Obviously, the drawings described below are only the implementation methods of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A flowchart of an information pollution assessment method in a large model provided in one embodiment of this application; Figure 2 A flowchart for constructing an input risk vector is provided as another embodiment of this application; Figure 3 A flowchart for constructing an output risk vector is provided as another embodiment of this application; Figure 4 A schematic diagram of an information pollution assessment system in a large model provided for the implementation of this application; Figure 5 This is a schematic diagram of the hardware structure of an electronic device provided for an embodiment of this application. Detailed Implementation

[0018] The present invention will now be described in detail with reference to the specific embodiments shown in the accompanying drawings. However, these embodiments do not limit the present invention, and any structural, methodological, or functional modifications made by those skilled in the art based on these embodiments are included within the scope of protection of the present invention.

[0019] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this application should have the ordinary meaning understood by those skilled in the art to which this application pertains. The terms "first," "second," and similar terms used in the embodiments of this application do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects.

[0020] With the widespread application of large-scale models in various scenarios, these models typically learn language patterns, knowledge structures, and reasoning models automatically from massive internet corpora. However, the sources of open internet data are complex. Once these low-quality or biased corpora are absorbed by the model, they will form biased representations or erroneous associations in the parameter space, manifesting in various anomalies during the reasoning and generation stages, thus weakening the reliability of the model's output. On the other hand, the output of large-scale models is reused and amplified multiple times in actual platforms. For example, generated content may be cached for historical record display, enter retrieval or recommendation systems as candidate results, be included in the index of retrieval augmented generation (RAG) systems, or even be used again as training samples during subsequent fine-tuning or continuous training, causing pollution to accumulate cyclically within the system and be difficult to eliminate.

[0021] Current content security technologies for large models mainly focus on the following directions, but overall they remain at the level of segmented and localized governance, and a unified evaluation framework that runs through the training data, output generation, and system reuse stages has not yet been formed.

[0022] Firstly, there are single-point detection techniques based on generated content. Much work focuses on fact-checking, illusion detection, harmful content identification, or bias detection of model output. Typical approaches involve using classification models, retrieval-based fact-checking systems, or rule filters to screen each piece of generated content individually to identify factual errors, logical contradictions, emotionally charged extreme expressions, or illegal and harmful elements. A common characteristic of these methods is that they only focus on the performance at the "output end," lacking simultaneous modeling of the "training corpus." They cannot characterize whether a particular error or bias originates from the training samples, in which corpus region it is more common, or whether the model amplifies or suppresses these risks during inference. Furthermore, single-point detection typically only targets a single generation action, making it difficult to cover the subsequent flow of generated content during caching, indexing, and retraining.

[0023] Secondly, data cleaning and auditing techniques based on training corpora: Some works attempt to deduplicatize, filter quality data, remove sensitive content, and conduct compliance audits on the data before model training to reduce noise and harmful samples in the training corpus. These methods can improve the overall purity of the training data to some extent, but they often remain at the level of overall data distribution or rule filtering, lacking refined risk quantification for individual training samples, and even more so, failing to establish a mapping relationship between "the risk characteristics of a certain training sample" and "the specific pollution manifestations of a certain generated content." In other words, existing data cleaning methods cannot answer "which type of abnormal sample in the training set does this erroneous output correspond to," nor can they measure whether a certain type of training pollution is reproduced or amplified on the output side.

[0024] Third, there are hallucination prevention techniques based on internal model behavior correction. These include methods such as supervised fine-tuning, reinforcement learning based on human feedback (RLHF), rejection strategies, counterfactual training, and model calibration to adjust the model's generation preferences and reduce the probability of obvious hallucinations and inappropriate outputs. These methods typically do not directly access the training corpus or monitor specific generated content; instead, they correct the model's behavior holistically by optimizing the objective or training process. While they can statistically improve the quality of model output, they cannot explicitly characterize the causal chain between "training-end contamination—output-end risk—system reuse," nor can they accurately attribute the cause of a single output, and they struggle to monitor whether contaminated content forms a loop in caches, retrieval indexes, or retraining data.

[0025] Fourth, there is currently a lack of a unified evaluation system covering the entire lifecycle of large models. Existing solutions from academia and industry do not yet meet the following requirements simultaneously: First, to simultaneously detect and quantitatively model contamination in both training corpora and generated content, rather than focusing on only one end; second, to construct multi-dimensional risk vectors for each training sample and each piece of generated content, covering at least key dimensions such as factuality, logicality, semantic quality, and sentiment bias; third, to establish a correspondence between training samples and generated content in a unified semantic space, thereby enabling cross-stage attribution of contamination and determining whether contamination is inherited from training data or amplified or newly generated during inference; fourth, to model the flow path of generated content in caches, search indexes, knowledge bases, and retraining data, characterizing the survival time, access frequency, and re-exposure risk of contamination within the system; and fifth, to construct a calculable global contamination index based on the above information, providing a quantitative basis for security strategies and governance actions.

[0026] In summary, existing technologies cannot answer key questions such as "whether a certain error or bias is significantly affected by a certain type of training corpus," "whether the model systematically amplifies training bias in certain semantic regions," and "whether a certain type of contamination content has been stored in the cache or index for a long time and has been called multiple times." Once contamination forms a cycle of reuse within the system, it is difficult to detect and effectively block it in a timely manner, resulting in a significant passivity and lag in contamination control.

[0027] To address the aforementioned problems, this application provides a method for assessing information pollution in a large model, such as... Figure 1 As shown, it includes the following steps: Step S100: Obtain the contaminated corpus for training the large model, and construct an input contamination vector for each piece of text in the contaminated corpus.

[0028] Specifically, the corpora in the polluted corpus are constructed into vector representations to form vector information. The polluted corpus is considered as a sample space, represented as (X, Y, P). train(), where X is the text input space. Y is the latent semantic label space, used to abstractly describe the attribute structure of the corpus, P train The empirical distribution of the training corpus is represented as D. train ={d1, d2, ..., d N}, where each sample d i All contain text content and its source, timestamps, credibility, and other metadata.

[0029] In one possible implementation of this application, step S100 includes: Step S110: Construct a multi-dimensional input pollution vector for each piece of text in the pollution corpus.

[0030] Specifically, in order to quantify the potential pollution of the corpus, multiple dimensions are constructed for each corpus to represent the quality of the corpus in different aspects.

[0031] In one possible implementation of this application, such as Figure 2 As shown, the multi-dimensional input contamination vector includes the dimension of factual distortion risk, the dimension of logical structure risk, the dimension of semantic quality risk, and the dimension of sentiment bias risk. These represent the quality of the corpus in terms of accuracy, logical structure, semantic expression, and sentiment bias, respectively.

[0032] Other possible implementation methods include multiple dimensions such as toxicity dimension or copyright conflict dimension.

[0033] In one possible implementation of this application, when constructing the fact distortion risk dimension, structured knowledge extraction is first performed on the corpus, transforming the text into a set of triples, which can be represented as triples(d i = {(h, r, t)}. Then, by aligning with a trusted fact base, a consistency score, denoted as s, is calculated for each triple. fact (h, r, t) ∈ [0, 1], the higher the consistency score of the triple, the more credible the training content is, and the lower the consistency score, the more likely it is to contain false facts or incorrect references.

[0034] The input contamination vector in the fact distortion risk dimension is calculated using the following formula: ; in, For the input or output contamination vector of the fact distortion risk dimension, d i For corpus, Let h, r, and t be the set of triples in the corpus, where h, r, and t are the head entity, relation, and tail entity in the triple, respectively. The consistency score for triples.

[0035] The above formula is the inverse vector of the consistency score, used to measure the degree of potential factual bias in the corpus.

[0036] In one possible implementation of this application, the text of the corpus is first split into a sequence of sentences, represented as d. i =[a1, a2, ..., a L Furthermore, it utilizes a natural language reasoning model to analyze the reasoning relationships between adjacent sentences, calculating the probability of contradiction between adjacent sentences, denoted as p. contra (s) k s k+1 If adjacent sentences have obvious semantic or logical conflicts, their probability of contradiction will increase significantly.

[0037] The input contamination vector in the logical structure risk dimension is calculated using the following formula: ; in, For the risk dimension of the logical result, either the input contamination vector or the output contamination vector. A sentence sequence, This represents the probability of contradiction between two adjacent sentences.

[0038] The numerical value of this formula reflects whether the internal logic of the text is coherent. A higher risk value indicates that there are structural problems such as logical jumps and contradictions in the training corpus.

[0039] In one possible implementation of this application, for each corpus, a syntactic structure score is calculated to reflect the grammatical completeness of the text, and a referential consistency score is also calculated to measure whether the pronoun reference is clear and unambiguous.

[0040] The input contamination vector in the semantic quality risk dimension is calculated using the following formula: ; in, q is the input or output contamination vector for the semantic quality risk dimension. syn To score the syntactic structure of the corpus, q coref The reference consistency score is given to the corpus.

[0041] The larger the result obtained by the formula, the greater the risk that the corpus has problems such as semantic ambiguity, unclear expression, or syntactic confusion.

[0042] In one possible implementation of this application, the emotional intensity and bias expression intensity of the corpus are calculated simultaneously, which are used to measure the emotional tendency and the degree of stereotype bias in the text of the corpus, respectively.

[0043] The input contamination vector in the sentiment bias wind dimension is calculated using the following formula: ; in, For the input or output contamination vector of the emotion bias towards the wind dimension, e(d) i b(d) represents the intensity of emotion. i () represents the intensity of biased expression.

[0044] Emotional bias risk reflects whether the corpus contains exaggerated expressions, emotional language, or potentially discriminatory content.

[0045] In summary, the multi-dimensional input contamination vector can be represented as: .

[0046] It should be noted that the contaminated corpus used for training belongs to the training corpus side of the large model. The contaminated corpus is considered the fundamental source of knowledge representation and reasoning patterns within the large model, and its contaminated content is believed to be written into the parameters of the large model in a specific form and activated during inference. Therefore, a multi-dimensional input contamination vector is constructed for each piece of text in the contaminated corpus. Specifically, at least four core risks are quantitatively characterized: factual distortion risk, logical structure risk, semantic quality risk, and sentiment bias risk. The input contamination vector for each risk dimension is calculated using a fixed detection method. Through this process, the corpus used for training is no longer just a large, unstructured collection of text, but is transformed into structured data with sample-level risk features, providing accurate source information for subsequent cross-link association analysis.

[0047] Step S200: Input the polluted corpus into the large model, use the large model to generate an output set, and construct an output pollution vector for each output content in the output set.

[0048] Specifically, the output content constitutes an output set, which is represented as: .

[0049] In one possible implementation of this application, step S200 includes: Step S210: Construct a multi-dimensional output contamination vector for each output item in the output set.

[0050] Specifically, in order to quantify the potential pollution of the output content, multiple dimensions are constructed for each piece of output content to represent the quality of the output content in different aspects.

[0051] In one possible implementation of this application, such as Figure 3 As shown, the multi-dimensional output contamination vector includes the dimension of fact distortion risk, the dimension of logical structure risk, the dimension of semantic quality risk, and the dimension of sentiment bias.

[0052] Specifically, to ensure consistency in the contamination system between the input and output content used for training, the same four-dimensional contamination vector is constructed. The calculation and judgment methods for the four dimensions of the output contamination vector are the same as those for the four dimensions of the input contamination vector described above, and will not be repeated here.

[0053] The output contamination vector, composed of four dimensions, is represented as follows: .

[0054] The output set belongs to the content generation side of the large model. The large model is considered as a family of conditional distributions learned from the training distribution. The output generated by the model when given input prompts is the explicit result of its internal knowledge and reasoning model. To comprehensively reflect the pollution manifestations in the generation behavior, a multi-dimensional output risk vector with the same structure as the training end is constructed for each generated content, quantifying dimensions such as factual illusion, logical incoherence, semantic ambiguity, and emotional misleading. The risk values ​​of each dimension are obtained through fixed testing methods, such as fact checking based on a credible fact base and verifiable citations, logical evaluation based on the integrity of the reasoning chain, semantic ambiguity evaluation based on the result of referential resolution, and emotional misleading judgment based on the emotional intensity threshold. In this way, the training corpus side and the generated content side achieve a dimensional and semantic correspondence in risk representation, forming an "input-side risk map" and an "output-side risk map".

[0055] Step S300: Based on the actual semantics of multiple corpora and multiple output contents, match the corpora and output contents whose actual semantics meet the similarity requirements, and calculate the inheritance pollution amount between each set of matched output contents and corpora.

[0056] Specifically, in order to determine whether the pollution in the output content comes from the training corpus and to determine whether the large model amplifies the pollution during the inference process, semantically identical corpus and output content are matched, and the amount of integrated pollution between the output content and the corpus is calculated.

[0057] In one possible implementation of this application, step S300 includes: Step S310: Using the same semantic encoding, multiple corpora and multiple output contents are mapped to the same semantic space.

[0058] Step S320: Calculate the similarity between multiple corpora and multiple output contents.

[0059] Step S330: Match the corpus with the highest similarity for each output content.

[0060] Specifically, a shared encoder is first used to map the training corpus and output content to the same semantic space. This shared semantic space allows for a direct connection between the two types of content through vector similarity. The shared encoder is represented as h(.), and the mapping of the corpus to the semantic space is represented as h(.). i =h(d i The output content is mapped to a semantic space representation as follows: Then, cosine similarity is used to measure the actual semantic closeness between the corpus and the output content. The higher the similarity, the greater the potential influence of the corpus on the output content. The similarity between the corpus and the output content is expressed as: Based on similarity, the top K most similar corpora are found for each output as candidate sources. This is represented as: .

[0061] In one possible implementation of this application, step S300 further includes: The weight of each output content matching the corpus is calculated using the following formula: ; Where, α j，i The weights of the corpus i that match the output content j, where λ represents the degree of concentration of the weight distribution, and h i The corpus is semantically encoded. The output content after semantic encoding, d i ∈N(y j () refers to corpora that meet the similarity requirements.

[0062] The weights of the semantic domain corpus are normalized. The larger the weight, the stronger the influence of the training corpus on the generated content.

[0063] Calculate the amount of inherited pollution using the following formula: ; in, To inherit the amount of pollution, This is the input contamination vector.

[0064] In summary, after constructing the dual-end risk representation in steps S100 and S200, the system further enters a unified semantic representation space. Each input corpus and each output content is mapped to a vector representation using the same encoder, and a set of semantic neighborhood samples is retrieved for each output content in the training space based on vector similarity. By weighted aggregation of the risk vectors of the neighborhood samples, the risk level of the output content "inherited from the training corpus" in each risk dimension is obtained.

[0065] Step S400: Calculate the pollution impact coefficient of the large model based on the output pollution vector and the inherited pollution amount.

[0066] In one possible implementation of this application, step S400 includes: Based on the output pollution vector and inherited pollution amount, calculate the pollution impact coefficient of the large model, including: The pollution impact coefficient is calculated using the following formula: ; Among them, A j The pollution impact coefficient, The output pollution vector is ε, which is a constant to prevent the denominator from being zero.

[0067] Specifically, the pollution impact coefficient is used to determine whether the large model amplifies pollution during the inference process. When the pollution impact coefficient A... j A value greater than 1 indicates that the large model amplifies contamination during the inference process. When the contamination impact coefficient A... j When the value is less than 1, it indicates that the large model will reduce the pollution during the inference process. The training goal of the large model is to continuously reduce the amplification of pollution during the inference process, so as to reduce the pollution. Therefore, the pollution influence coefficient of the large model will be calculated multiple times during the training process to evaluate the degree of information pollution of the large model.

[0068] It should also be noted that the output content inherits the risk level from the training corpus across various risk dimensions, and compares it with the actual observed risks at the generation end to calculate the contamination impact coefficient. Based on the contamination impact coefficient, it is determined whether the large model amplifies or suppresses contamination. Through this cross-stage contamination association modeling process, this invention can determine at the sample level whether a certain type of contamination is mainly integrated from the self-training data and amplified during inference, or whether it is a new bias generated in a region with low risk in the training data, thereby achieving interpretable identification of the causes of contamination.

[0069] Step S500: Construct nodes in the application system of the large model, and use the nodes to record the node risk value of the output content in the application system.

[0070] Specifically, the output of large models may be cached, indexed, recorded in a knowledge base, or used for training within the application system, potentially creating a pollution cycle. A better approach is to model the content within the application system as a set of nodes. Each node records the node risk value of the output content at that node, and the node risk values ​​of all nodes constitute the node risk value of the output content within the application system.

[0071] In one possible implementation of this application, step S500 includes: The system uses nodes to record the dynamic behavior information of the output content, including the first write time, the last access time, and the number of calls.

[0072] Calculate the node risk value using the following formula: ; ; in, The node risk value is the value calculated for the dimension in question. Let m be the node, and m be the node ordinal number. The most recent access time, The first write time is k, and k is the dimension. The pollution vector is in dimension 1. For node lifespan, This represents the number of times the function is called.

[0073] Node inventory cycle T m The longer the cycle, the longer the contaminated content persists in the application system, and the higher the risk of impacting subsequent content.

[0074] In this step, we focus on the lifecycle evolution of contamination in real application systems and construct a contamination lifecycle tracking mechanism. This mechanism uniformly models the writing and access behavior of output content in the caching system, search index, knowledge base, and subsequent retraining data. Each internal storage entry is regarded as a node with contamination characteristics, storage time, and access frequency. Nodes with "high contamination + long survival + high frequency of calls" are marked and their risks are amplified, thereby depicting the flow path of contamination within the system and the risk of re-exposure.

[0075] Step S600: Calculate the global pollution risk index based on multiple input pollution vectors, multiple output pollution vectors, and multiple node risk values. The pollution impact coefficient and the global pollution risk index characterize the degree of information pollution in the large model.

[0076] Specifically, to achieve a unified quantification of the overall contamination risk of the large model, the risks at the training end, generation end, and system end are integrated. The input contamination vector represents the training end risk, the output contamination vector represents the generation end risk, and the node risk value represents the system end risk. This provides the platform with a configurable and computable global risk metric for a full lifecycle information contamination assessment, making the generation, propagation, and amplification of contamination quantifiable, traceable, and explainable.

[0077] In one possible implementation of this application, step S600 includes: Step S610: Calculate the mean input pollution of multiple input pollution vectors of multiple corpora, calculate the mean output pollution of multiple output pollution vectors of multiple output contents, and calculate the mean system node risk of multiple node values ​​of multiple output contents.

[0078] Step S620: Calculate the comprehensive risk index according to the following formula: ; Among them, R k As a comprehensive risk index, To input the average pollution level, To output the average pollution level, β1, β2, and β3 are the system node averages, β3 are the scene coefficients, and k is the dimension.

[0079] First, calculate the comprehensive risk value for one dimension, which combines the average input contamination value of multiple input contamination vectors from multiple corpora in one dimension, the average output contamination value of multiple output contamination vectors, and the system node average risk value of multiple nodes. The scenario coefficient can be flexibly adjusted according to the business scenario.

[0080] Step S630: Calculate the global pollution risk index according to the following formula: ; Where R global The global pollution risk index is represented by n, where n is the total number of dimensions, and w is the total number of dimensions. k The weights are for dimension k.

[0081] Specifically, a global pollution index is obtained by combining multiple dimensions of comprehensive risk index. This index can be used to automatically trigger multi-level governance strategies, including content blocking, citation replacement, fact correction, model correction, and training corpus cleaning, thereby achieving closed-loop control of pollution from "detection to location to tracking to governance".

[0082] In one feasible approach, the risk of factual distortion, logical structure, semantic quality, and sentiment bias are considered together to calculate... The global pollution risk index is then obtained. In other feasible implementations, a toxicity dimension and a copyright conflict dimension are added to the above four dimensions before calculation. The overall pollution risk index was obtained.

[0083] In one possible implementation of this application, the step S600 includes the following: Step S710: Set a pollution risk index threshold. When the global pollution risk index is greater than the pollution risk index threshold, the output content is intercepted.

[0084] Step S720: Analyze the knowledge reference link of the intercepted output content, query the input content corresponding to the output content, and correct the input content.

[0085] Specifically, when the pollution risk index of a certain dimension or the overall pollution risk index exceeds the pollution risk index threshold, the platform can automatically enter the governance mode: First, content-level security interception is carried out on the output side, and risky content is degraded, partially deleted, replaced with templated answers, or forcibly rejected to block the spread of pollution. Subsequently, in the knowledge citation chain, consistency verification is performed on the factual items, source citations, and knowledge fragments involved in the output content, and untrusted citations are replaced with verified factual data or standardized knowledge, thereby restoring the credibility of the output content.

[0086] In one possible implementation of this application, the method further includes the following after step S600: Step S810: In response to the fact that the numerical representation of the pollution impact coefficient of the large model amplifies the pollution content of the corpus, the reasoning logic of the large model is corrected.

[0087] Specifically, it is also necessary to further trigger the correction mechanisms of the parameter and inference layers of the large model. When the contamination impact coefficient indicates that the model exhibits continuous amplification behavior in a certain dimension, lightweight correction modules such as local fine-tuning, counterfactual training, and negative example reinforcement can be automatically activated, allowing the large model to restore inference stability without requiring complete retraining. Simultaneously, a root-cause investigation will be conducted on contaminated content stored in the cache, index, vector database, and the RAG index generated by retrieval enhancement. High-risk nodes will undergo content removal, vector replacement, index reconstruction, or data cleaning to prevent the recycling of contamination within the system.

[0088] It should be noted that the specific explanation for the amplification behavior of the large model of the pollution impact coefficient for the pollution content of the corpus is as described in step S400, that is, the calculation result of the formula in step S400 is greater than 1, which will not be repeated here.

[0089] In one specific implementation, a feedback-based dynamic optimization mechanism is initiated after step S720 or S810. After the governance measures are implemented, the relevant content is re-verified through the training end and the pollution detector that generated the daunt, and the pollution impact coefficient and global pollution risk index are recalculated. If the risk is effectively reduced, the pollution status map of the system is updated; if the risk is still too high, the intensity of the governance strategy is automatically adjusted, such as increasing the interception level, tightening the fact verification threshold, or expanding the data range for reverse cleaning. Through this continuous feedback loop, the pollution governance process can be ensured to have adaptability and long-term stability, so that the model can maintain a high degree of controllable information credibility and content security even under long-term operation, frequent iteration, or high-concurrency usage scenarios.

[0090] In one of the possible implementations of this application, a method for assessing information pollution in a large model is provided, the beneficial effects of which include: A dual-link pollution modeling system, encompassing both training and generation, is constructed. This system quantifies the quality of training data, model inference behavior, and output content risk using a unified risk vector, achieving a complete correlation between pollution sources and final output. Through this mechanism, the platform can accurately pinpoint the source of pollution, identify its formation mechanism, and implement differentiated governance based on its causes, thus improving the controllability of model pollution management. Furthermore, by employing a unified semantic vector space, semantic neighborhood retrieval, and weighted attribution mechanisms, each output piece of content can establish a computable association with potential sources in the training corpus. By calculating the pollution inheritance amount and amplification coefficient, the system distinguishes between "inherited pollution from the training corpus," "new pollution generated by inference," and "model amplification behavior," thereby achieving more refined pollution source identification and targeted governance.

[0091] This method tracks and assesses the cumulative risk of contaminated content within the system by recording node lifespan, access frequency, and content contamination intensity. It can determine whether contaminated content is repeatedly invoked, remains for extended periods, or becomes a hidden source for subsequent tasks, enabling continuous monitoring of the internal contamination propagation chain. This prevents contamination from forming a closed loop within the system, significantly improving platform security and long-term model stability. A weighted contamination risk index is formed by weighting and integrating training-end risks, generation-end risks, and internal system flow risks. This index can be dynamically adjusted based on business scenarios, compliance requirements, and model versions, providing quantifiable and actionable decision-making support for content filtering, output monitoring, and model optimization strategies, transforming contamination management from experience-based to engineering-based. Furthermore, a four-stage closed-loop mechanism of detection, association, tracking, and governance ensures that contamination is not only identifiable and explainable but also allows for real-time intervention and automatic repair, achieving system-level contamination control throughout the model's lifecycle and significantly improving the platform's long-term stability and security.

[0092] Another embodiment of this application discloses an information pollution assessment system in a large model, such as Figure 4 As shown, it includes: The input risk module is used to obtain a polluted corpus for training large models and to construct an input pollution vector for each piece of text in the polluted corpus. The output risk module is used to input the pollution corpus into the large model, use the large model to generate an output set, and construct an output pollution vector for each output content in the output set; The first matching module is used to match the actual semantics of multiple corpora and multiple output contents with the corpora whose actual semantics meet the similarity requirements, and to calculate the amount of inheritance pollution between the output contents and the corpora after each matching. The first calculation module is used to calculate the pollution impact coefficient of the large model based on the output pollution vector and the inherited pollution amount; The system risk module constructs nodes in the application system of the large model and uses these nodes to record the node risk value of the output content in the application system. The evaluation module is used to calculate the global pollution risk index based on multiple input pollution vectors, multiple output pollution vectors, and multiple node risk values. The pollution impact coefficient and the global pollution risk index characterize the degree of information pollution in the large model.

[0093] In one possible implementation of this application, the first matching module includes: The mapping module is used to map multiple corpora and multiple output contents to the same semantic space using the same semantic code; The second calculation module is used to calculate the similarity between multiple corpora and multiple output contents; The second matching module is used to match the corpus with the highest similarity for each output content.

[0094] In one possible implementation of this application, the first matching module further includes: The third calculation module calculates the weight of the corpus matching each output content according to the following formula: ; Where, α j，i The weights of the corpus i that match the output content j, where λ represents the degree of concentration of the weight distribution, and h i The corpus is semantically encoded. The output content after semantic encoding, d i ∈N(y j () refers to corpora that meet the similarity requirements; Calculate the amount of inherited pollution using the following formula: ; in, To inherit the amount of pollution, The input pollution vector; Based on the output pollution vector and inherited pollution amount, calculate the pollution impact coefficient of the large model, including: The pollution impact coefficient is calculated using the following formula: ; Among them, A j The pollution impact coefficient, The output pollution vector is ε, which is a constant to prevent the denominator from being zero.

[0095] In one possible implementation of this application, the system risk module includes: The recording module is used to record the dynamic behavior information of the output content in the system using nodes. The dynamic behavior information includes the first write time, the last access time, and the number of calls. The fourth calculation module is used to calculate the node risk value according to the following formula: ; ; Specifically, The node risk value is the value calculated for the dimension in question. Let m be the node, and m be the node ordinal number. The most recent access time, The first write time is k, and k is the dimension. The pollution vector is in dimension 1. For node lifespan, This represents the number of times the function is called.

[0096] In one possible implementation of this application, the evaluation module includes: The fifth calculation module calculates the input pollution mean of multiple input pollution vectors of multiple corpora, the output pollution mean of multiple output pollution vectors of multiple output contents, and the system node mean of multiple node risk values ​​of multiple output contents. The comprehensive risk index is calculated using the following formula: ; Among them, R k As a comprehensive risk index, To input the average pollution level, To output the average pollution level, β1, β2, and β3 are the system node averages, β3 are the scene coefficients, and k is the dimension. The global pollution risk index is calculated using the following formula: ; Where R global The global pollution risk index is represented by n, where n is the total number of dimensions, and w is the total number of dimensions. k The weights are for dimension k.

[0097] In one possible implementation of this application, the input risk module includes: The first construction module is used to construct a multi-dimensional input pollution vector for each piece of text in the pollution corpus.

[0098] The risk output module includes: The second building module is used to construct a multi-dimensional output contamination vector for each output item in the output set.

[0099] In one possible implementation of this application, the input pollution vectors and output pollution vectors of the first construction module in multiple dimensions all include the dimension of fact distortion risk, the dimension of logical structure risk, the dimension of semantic quality risk, and the dimension of sentiment bias. The second building module's multi-dimensional input pollution vector and multi-dimensional output pollution vector both include the dimension of fact distortion risk, the dimension of logical structure risk, the dimension of semantic quality risk, and the dimension of sentiment bias.

[0100] The input or output contamination vector in the fact distortion risk dimension is calculated using the following formula: ; in, For the input or output contamination vector of the fact distortion risk dimension, d i For corpus, Let h, r, and t be the set of triples in the corpus, where h, r, and t are the head entity, relation, and tail entity in the triple, respectively. The consistency score for triples; The corpus is split into sentence sequences, and the input pollution vector or output pollution vector in the logical structure risk dimension is calculated according to the following formula: ; ; in, For the risk dimension of the logical result, either the input contamination vector or the output contamination vector. A sentence sequence, The probability of a contradiction between two adjacent sentences; The input or output contamination vector in the semantic quality risk dimension is calculated using the following formula: ; in, q is the input or output contamination vector for the semantic quality risk dimension. syn To score the syntactic structure of the corpus, q coref The reference consistency score is given to the corpus. The input or output contamination vector in the sentiment bias wind dimension is calculated using the following formula: ; in, For the input or output contamination vector of the emotion bias towards the wind dimension, e(d) i b(d) represents the intensity of emotion. i () represents the intensity of biased expression.

[0101] The implementation method of this application also includes: The interception module is used to set a pollution risk index threshold. When the global pollution risk index is greater than the pollution risk index threshold, the output content is intercepted. The correction module is used to analyze the knowledge reference links of the intercepted output content, query the corresponding input content, and correct the input content.

[0102] The implementation method of this application also includes: The correction module is used to correct the inference logic of the large model in response to the amplification behavior of the pollution impact coefficient on the pollution content of the corpus.

[0103] Figure 5 This diagram illustrates a more specific hardware structure of an electronic device provided in this embodiment. The device may include: a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, memory 1020, input / output interface 1030, and communication interface 1040 are interconnected internally via the bus 1050.

[0104] The processor 1010 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.

[0105] The memory 1020 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1020 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program code is stored in the memory 1020 and is called and executed by the processor 1010.

[0106] The input / output interface 1030 is used to connect input / output modules to realize information input and output. The input / output modules can be configured as components in the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touch screens, microphones, various sensors, etc., and output devices may include displays, speakers, vibrators, indicator lights, etc.

[0107] The communication interface 1040 is used to connect the communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, radio (shortwave / ultra-shortwave) communication, satellite communication, data link communication, etc.).

[0108] Bus 1050 includes pathways for transmitting information between various components of the device, such as processor 1010, memory 1020, input / output interface 1030, and communication interface 1040.

[0109] It should be noted that although the above-described device only shows the processor 1010, memory 1020, input / output interface 1030, communication interface 1040, and bus 1050, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the embodiments described in this specification, and need not include all the components shown in the figures.

[0110] The device described above is used to implement the information pollution assessment method in the corresponding large model in any of the foregoing embodiments, and has the beneficial effects of the corresponding method implementation, which will not be elaborated here.

[0111] Based on the same inventive concept, corresponding to any of the above-described embodiments, this application also provides a non-transitory computer-readable storage medium storing computer instructions for causing the computer to execute the information pollution assessment method in the large model as described in any of the above embodiments.

[0112] The computer-readable medium in this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0113] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to execute the information pollution assessment method in the large model as described in any of the above embodiments, and have the beneficial effects of the corresponding method implementation, which will not be repeated here.

[0114] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of this application (including the claims) is limited to these examples; this manner of description is merely for clarity, and those skilled in the art should consider the specification as a whole. Within the framework of this application, the above embodiments or the technical features of different embodiments can also be appropriately combined, the steps can be implemented in any order, and there are many other variations of different aspects of the embodiments of this application as described above, which are not provided in the details for the sake of brevity.

[0115] Additionally, to simplify the description and discussion, and to avoid obscuring the embodiments of this application, the well-known power / ground connections to integrated circuit (IC) chips and other components may or may not be shown in the provided drawings. Furthermore, the apparatus may be shown in block diagram form to avoid obscuring the embodiments of this application, and this also takes into account the fact that the details of the implementation of these block diagram apparatuses are highly dependent on the platform on which the embodiments of this application will be implemented (i.e., these details should be entirely within the understanding of those skilled in the art). While specific details (e.g., circuits) are set forth to describe exemplary embodiments of this application, it will be apparent to those skilled in the art that the embodiments of this application can be implemented without these specific details or with variations thereof. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0116] Although this application has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may use the embodiments discussed.

[0117] The embodiments described herein are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made without departing from the spirit and principles of the embodiments described herein should be included within the protection scope of this application.

Claims

1. A method for assessing information pollution in a large-scale model, characterized in that, The method includes: Obtain a polluted corpus for training the large model, and construct an input pollutant vector for each piece of text in the polluted corpus; The polluted corpus is input into the large model, and the large model is used to generate an output set. An output pollution vector is constructed for each output content in the output set. Based on the actual semantics of multiple corpora and multiple output contents, match the corpora and output contents whose actual semantics meet the similarity requirements, and calculate the inheritance pollution amount between each set of matched output contents and the corpora; Based on the output pollution vector and the inherited pollution amount, calculate the pollution impact coefficient of the large model; Nodes are constructed in the application system of the large model, and the nodes are used to record the node risk value of the output content in the application system. Based on the multiple input contamination vectors, the multiple output contamination vectors, and the multiple node risk values, a global contamination risk index is calculated. The contamination impact coefficient and the global contamination risk index characterize the degree of information contamination in the large model.

2. The information pollution assessment method in a large model according to claim 1, characterized in that, The step of matching the actual semantics of multiple corpora and multiple output contents with the corpora and output contents that meet the similarity requirements includes: Using the same semantic encoding, multiple corpora and multiple output contents are mapped to the same semantic space; Calculate the similarity between the multiple corpora and the multiple output contents; For each output content, match the corpus with the highest similarity.

3. The method for assessing information pollution in a large model according to claim 2, characterized in that, The calculation of the inheritance pollution amount between the output content after each matching and the corpus includes: The weight of each output content matching the corpus is calculated according to the following formula: ； Where, α j，i The weights of the corpus i that match the output content j, where λ represents the degree of concentration of the weight distribution, and h i The corpus after semantic encoding, The output content after semantic encoding, d i ∈N(y j The corpus that meets the similarity requirements; The amount of inherited contamination is calculated using the following formula: ； in, The amount of inherited contamination. The input contamination vector; The step of calculating the pollution impact coefficient of the large model based on the output pollution vector and the inherited pollution amount includes: The pollution impact coefficient is calculated using the following formula: ； Among them, A j The pollution impact coefficient is... Let ε be the output contamination vector, and let ε be a constant to prevent the denominator from being zero.

4. The method for assessing information pollution in a large model according to claim 1, characterized in that, The process of constructing nodes in the application system of the large model, and using these nodes to record the node risk values ​​of the output content in the application system, includes: The nodes are used to record the dynamic behavior information of the output content in the system, including the first write time, the most recent access time, and the number of calls. Calculate the node risk value using the following formula: ； ； Specifically, The node risk value is the value calculated for the dimension in question. Let m be the node, and m be the node ordinal number. The most recent access time, The first write time is k, where k is the dimension. The pollution vector is in dimension 1. For node lifespan, This represents the number of times the function is called.

5. The method for assessing information pollution in a large model according to claim 1, characterized in that, The step of calculating the global contamination risk index based on multiple input contamination vectors, multiple output contamination vectors, and multiple node risk values ​​includes: Calculate the mean input contamination of multiple input contamination vectors of multiple corpora, calculate the mean output contamination of multiple output contamination vectors of multiple output contents, and calculate the mean system node risk of multiple node values ​​of multiple output contents; The comprehensive risk index is calculated using the following formula: ； Among them, R k This refers to the comprehensive risk index. The input pollution mean, The average value of the output pollution. β1, β2, and β3 are the average values ​​of the system nodes, β3 are the scene coefficients, and k is the dimension. The global pollution risk index is calculated using the following formula: ； Where R global Let n be the global pollution risk index, and w be the total number of dimensions. k The weights are for dimension k.

6. The method for assessing information pollution in a large model according to claim 1, characterized in that, The step of constructing an input pollution vector for each piece of text in the pollution corpus includes: For each piece of text in the polluted corpus, construct a multi-dimensional input pollution vector; The step of constructing an output contamination vector for each output item in the output set includes: For each output item in the output set, construct a multi-dimensional output contamination vector.

7. The method for assessing information pollution in a large model according to claim 6, characterized in that, The input contamination vector and the output contamination vector, which are in multiple dimensions, both include the dimension of fact distortion risk, the dimension of logical structure risk, the dimension of semantic quality risk, and the dimension of sentiment bias. The input contamination vector or the output contamination vector in the fact distortion risk dimension is calculated according to the following formula: ； in, For the input contamination vector or the output contamination vector of the fact distortion risk dimension, d i For the corpus, Let h, r, and t be the set of triples in the corpus, where h, r, and t are the head entity, relation, and tail entity in the triple, respectively. The consistency score for triples; The corpus is split into sentence sequences, and the input contamination vector or the output contamination vector in the logical structure risk dimension is calculated according to the following formula: ； ； in, The input contamination vector or the output contamination vector is the risk dimension of the logical result. For the sentence sequence, The probability of a contradiction between two adjacent sentences; The input contamination vector or the output contamination vector in the semantic quality risk dimension is calculated according to the following formula: ； in, For the input contamination vector or the output contamination vector of the semantic quality risk dimension, q syn The syntactic structure score of the corpus is given by q. coref The reference consistency score is assigned to the corpus. The input contamination vector or the output contamination vector in the emotional bias wind dimension is calculated according to the following formula: ； in, For the input contamination vector or the output contamination vector of the emotional bias wind dimension, e(d) i ) represents the intensity of emotion, b(d) i () represents the intensity of biased expression.

8. The method for assessing information pollution in a large model according to claim 1, characterized in that, After characterizing the information pollution level of the large model using the pollution impact coefficient and the global pollution risk index, the following is included: Set a pollution risk index threshold; when the global pollution risk index is greater than the pollution risk index threshold, the output content will be blocked. Analyze the knowledge reference links of the intercepted output content, query the input content corresponding to the output content, and correct the input content.

9. The method for assessing information pollution in a large model according to claim 1, characterized in that, After characterizing the information pollution level of the large model using the pollution impact coefficient and the global pollution risk index, the following is included: In response to the numerical characterization of the pollution impact coefficient indicating that the large model amplifies the pollution content of the corpus, the inference logic of the large model is corrected.

10. An information pollution assessment system in a large model, characterized in that, The system includes: The input risk module is used to obtain a pollution corpus for training the large model and to construct an input pollution vector for each piece of text in the pollution corpus. The output risk module is used to input the pollution corpus into the large model, use the large model to generate an output set, and construct an output pollution vector for each output content in the output set. The first matching module is used to match the actual semantics of the multiple corpora and the multiple output contents with the corpora whose actual semantics meet the similarity requirements, and to calculate the amount of inheritance pollution between the output contents and the corpora after each matching; The first calculation module is used to calculate the pollution impact coefficient of the large model based on the output pollution vector and the inherited pollution amount; The system risk module constructs nodes in the application system of the large model and uses these nodes to record the node risk values ​​of the output content in the application system. The evaluation module is used to calculate a global pollution risk index based on multiple input pollution vectors, multiple output pollution vectors, and multiple node risk values. The pollution impact coefficient and the global pollution risk index characterize the degree of information pollution in the large model.

Images