A multi-dimensional retrieval enhanced generation-based large model event analysis enhancement method

Abstract

The application provides a multi-dimensional retrieval enhanced large model event analysis enhancement method, and relates to the cross technical field of artificial intelligence and information analysis. After obtaining the information to be retrieved, on the one hand, the network retrieval result is queried on the network, the picture-text consistency evaluation result and the semantic authenticity evaluation result are obtained for the network retrieval result, and the first corpus data is integrated by using the network retrieval result, the picture-text consistency evaluation result and the semantic authenticity evaluation result. On the other hand, the information to be retrieved is generated into a retrieval vector in multiple dimensions, the knowledge retrieval result of each dimension is retrieved in the corresponding preset knowledge database, and the second corpus data is integrated by using the knowledge retrieval results of multiple dimensions. The enhanced corpus obtained by merging the first corpus data and the second corpus data is used as semantic data for the large model to perform analysis. When the large model uses the semantic data provided by the application scheme to perform an analysis task, a more reliable analysis result can be obtained.

Description

Technical Field

[0001] This application relates to the interdisciplinary field of artificial intelligence and information analysis, specifically to a method for enhancing event analysis of large models based on multi-dimensional retrieval enhancement. Background Technology

[0002] Large Language Models (LLMs, referred to as large models in this paper) are widely used, and their ability to obtain semantic data by relying on external retrieval-Augmented Generation (RAG) is a key support.

[0003] However, the standard RAG architecture faces serious credibility challenges: On the one hand, its online search retrieval sources are open networks with inconsistent content quality, easily introducing misleading or biased information. When this information is used as semantic data input into a large model, it can lead to "illusionary" outputs containing factual errors or logical contradictions, weakening the reliability of the analysis conclusions. On the other hand, current large model outputs generally adopt a single-channel RAG framework driven by a general embedding model: that is, using a single text encoder (such as S-BERT) to vectorize the entire knowledge base, and then using the most semantically similar document fragments retrieved from the user query as context input to the large model. This method has the following drawbacks:

[0004] First, there is a lack of a credibility assessment mechanism for the retrieved content. Existing RAG systems typically assume that the search results are either entirely credible or entirely unreliable: when deemed entirely credible, they directly incorporate the results as semantic data into the prompts for use by the large model, without any form of falsification processing of the original information. When false information is retrieved, the large model treats it as true evidence for reasoning, thus amplifying errors. When deemed entirely unreliable, even if some information is true (e.g., the subject or time is correct but the event details are fabricated), RAG cannot identify and retain credible fragments, but instead directly rejects all information. This approach leads to an "all or nothing" information utilization pattern, which may introduce erroneous content or, due to excessive caution, ignore potentially valid clues.

[0005] Second, the retrieval dimensions are too limited to support complex information reasoning. Some types of event analysis require the integration of multi-dimensional heterogeneous information, such as the behavioral characteristics of the subjects, the historical context of events, specific rule-based approaches, and expert subjective opinions. However, general embedding models can only capture surface semantic similarity and cannot distinguish the aforementioned orthogonal dimensions, resulting in a narrow analytical perspective and one-sided conclusions.

[0006] Therefore, there is an urgent need for a new type of large-scale model information analysis credibility enhancement scheme to solve the above problems. Summary of the Invention

[0007] The technical problem this application aims to solve is that existing retrieval enhancement generation technologies have poor anti-counterfeiting capabilities and low credibility of large model information analysis due to single retrieval dimensions. Therefore, it provides a method for enhancing large model event analysis based on multi-dimensional retrieval enhancement generation.

[0008] Firstly, the technical solution of this application provides a method for enhancing event analysis of large models based on multi-dimensional retrieval enhancement, including:

[0009] S100: Obtain the information to be retrieved;

[0010] S200: Based on the information to be retrieved, several online search results are obtained online; each online search result includes text and image; for each online search result, the text-image consistency evaluation result and the semantic authenticity evaluation result of the text are calculated; the first corpus data is obtained by integrating each online search result and its corresponding text-image consistency evaluation result and semantic authenticity evaluation result.

[0011] S300: The information to be retrieved is input into a multi-dimensional preset embedding model to obtain a multi-dimensional retrieval vector. The multiple dimensions include at least two of the following: subject feature dimension, recent event dimension, historical event dimension, and subjective argument dimension. The retrieval vector of each dimension is used to query the corresponding dimension's preset knowledge database to obtain the knowledge retrieval result corresponding to each dimension. The knowledge retrieval results of multiple dimensions are integrated to obtain the second corpus data.

[0012] S400: The first corpus data and the second corpus data are merged to obtain the enhanced corpus, which is used as semantic data for analysis by the large model.

[0013] Preferably, the large model event analysis enhancement method based on multi-dimensional retrieval enhancement specifically includes step S200 as follows:

[0014] S201: Extract keywords from the information to be retrieved, and search the internet to obtain several online search results related to the keywords;

[0015] S202: For each web search result, extract the text and images; wherein the text includes the title, body text, source, and publication time; and the images include accompanying pictures and video frame images.

[0016] S203: For each web search result, the text is processed into standardized text, the image is processed into standardized image, and the standardized text and the standardized image are merged to obtain a standardized image-text combination;

[0017] S204: Input the standardized image-text combination and the standardized text corresponding to each network retrieval result into the trained image-text consistency detection model and semantic topic authenticity detection model in parallel. The image-text consistency detection model outputs the image-text consistency evaluation result, and the semantic topic authenticity detection model outputs the semantic authenticity evaluation result.

[0018] S205: Arrange each of the network retrieval results and its corresponding image-text consistency evaluation results and semantic authenticity evaluation results in order to obtain the first corpus data.

[0019] Preferably, in the large model event analysis enhancement method based on multi-dimensional retrieval enhancement, in step S204, the image-text consistency authentication model is obtained in the following way:

[0020] S2041: Select the CLIP architecture as the first initial model, wherein the CLIP architecture includes a text encoder, an image encoder, and a cross-modal fusion unit; wherein:

[0021] The text encoder is used to encode the standardized text into a q-dimensional semantic vector: , This represents the text encoding function, where T represents the standardized text. Let R denote the q-dimensional embedding space, and let R denote the spatial domain;

[0022] The image encoder is used to encode the normalized image into a q-dimensional visual vector: , I represents the image coding function, and I represents the normalized image.

[0023] The cross-modal fusion engine is used to weight and fuse semantic vector features and visual vector features and output a fused feature vector: , h is the cross-attention operation function. T For text query features, h I Image key-value features;

[0024] S2042: A first classification layer is set after the first initial model. The first classification layer is pre-trained using text-image pairs and consistency probability labels of each text-image pair from a specific dataset. Training is completed when the first difference between the consistency probability prediction result calculated by the first classification layer and the consistency probability label satisfies the convergence condition, wherein: C mm For the consistency probability prediction result and C mm ∈[0,1], It is the sigmoid activation function. ∈R1×q For trainable weights, ∈R represents a bias term;

[0025] The first difference is calculated using the cross-entropy loss function: ; Let be the consistency probability label for the i-th text-image pair, with a value of 1 indicating consistency and a value of 0 indicating inconsistency; N is the total number of text-image pairs. The consistency probability prediction result for the i-th text-image pair;

[0026] S2043: The image-text consistency authentication model is obtained by encapsulating the first initial model and the first classification layer; wherein, after receiving the fused feature vector, the consistency probability prediction result output by the first classification layer is used as the image-text consistency evaluation result.

[0027] Preferably, in the large model event analysis enhancement method based on multi-dimensional retrieval enhancement, in step S204, the semantic topic authenticity authentication model is obtained in the following way:

[0028] S2044: Select the DeBERTa-v3 architecture as the second initial model, which includes a text feature extractor; the text feature extractor is used to: obtain a context-aware sequence of the normalized text: Where S is the sequence length and d is the hidden layer dimension; the global semantics of the standardized text are aggregated using a weighted pooling algorithm: ;in, This represents the perceptual sequence of the i-th standardized text. This represents the perceptual sequence of the j-th standardized text. Let w represent the weighting coefficient of the i-th standardized text, where w ∈ R. d For learning parameters;

[0029] S2045: A second classification layer is set after the second initial model. The second classification layer is pre-trained using text from a specific dataset and the real labels of each text. Training is completed when the second difference between the real prediction result calculated by the second classification layer and the real label satisfies the convergence condition, wherein: ； p text For the true prediction result and p text ∈[0,1];

[0030] The second difference is calculated using the cross-entropy loss function: , Let be the authenticity label for the i-th text, with a value of 1 indicating authenticity and a value of 0 indicating inauthenticity; M is the total number of texts. The result of the authenticity prediction for the i-th text;

[0031] S2046: The semantic topic authenticity detection model is obtained by encapsulating the second initial model and the second classification layer; wherein, after receiving the standardized text, the second classification layer outputs the authenticity prediction result as the semantic authenticity evaluation result.

[0032] Preferably, in the large model event analysis enhancement method based on multi-dimensional retrieval enhancement, in step S300, the pre-embedded model for each dimension is obtained in the following way:

[0033] S301: Select a corpus corresponding to the current dimension, and extract triples that conform to the semantics of the dimension from the corpus. The triples include anchor points, positive samples, and negative samples.

[0034] S302: Select the S-BERT architecture as the initial embedding model, wherein the S-BERT architecture includes a triplet marginal loss function; train the initial embedding model using the triplets, and complete the training when the triplet marginal loss result calculated by the initial embedding model satisfies the convergence condition, wherein the triplet marginal loss function is:

[0035] ;

[0036] in, This represents the anchor vector of the i-th triplet; This represents the positive sample vector of the i-th triplet. Let K represent the negative sample vector of the i-th triplet; K represents the total number of sample vectors. Indicates the marginal threshold; This represents the Euclidean distance regularization term; Indicates the regularization weight;

[0037] S303: Use the initial embedding model that has been trained as a preset embedding model;

[0038] The pre-set embedding model for each dimension is used to convert the information to be retrieved into a retrieval vector for that dimension, and also to convert the knowledge information base for that dimension into a vector base.

[0039] Preferably, in the large model event analysis enhancement method based on multi-dimensional retrieval enhancement, in step S300:

[0040] The pre-built knowledge database for each dimension includes a knowledge information database, a vector database, and mapping relationships for that dimension, wherein:

[0041] The knowledge information base includes knowledge information under this dimension;

[0042] Each piece of knowledge information is processed using the pre-set embedding model to obtain a knowledge vector, and the knowledge vectors form the vector library.

[0043] The mapping relationship is the mapping relationship between each knowledge vector in the vector library and the corresponding knowledge information in the knowledge information library.

[0044] Preferably, in the large model event analysis enhancement method based on multi-dimensional retrieval enhancement, step S300 involves using the retrieval vector of each dimension to query the corresponding dimension's pre-set knowledge database to obtain the knowledge retrieval result for each dimension, and integrating the knowledge retrieval results from multiple dimensions to obtain the second corpus data, including:

[0045] S304: Query the vector library of the preset knowledge database for the knowledge vector that has the highest similarity to the retrieval vector as the target vector;

[0046] S305: Based on the mapping relationship, search the knowledge information in the knowledge information base that corresponds to the target vector as the target knowledge information;

[0047] S306: The knowledge retrieval result for that dimension is obtained by combining the dimension and the target knowledge information.

[0048] Secondly, the present application provides a computer-readable storage medium storing program information. After reading the program information, the computer executes the steps of the large model event analysis enhancement method based on multi-dimensional retrieval enhancement generation as described in any one of the first aspects.

[0049] Thirdly, the present application provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the large model event analysis enhancement method based on multi-dimensional retrieval enhancement generation as described in any of the first aspects.

[0050] Fourthly, the present application provides an electronic device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the large model event analysis enhancement method based on multi-dimensional retrieval enhancement generation as described in any of the first aspects.

[0051] The technical solution provided in this application has the following technical effects compared with the prior art:

[0052] The event analysis enhancement method for large-scale models based on multi-dimensional retrieval enhancement provided in this application performs two-stage retrieval processing after obtaining the information to be retrieved. First, it queries online search results and evaluates the consistency between text and images and the semantic authenticity of these results, obtaining text-image consistency evaluation results and semantic authenticity evaluation results. These three evaluation results are then integrated to obtain the first corpus data. Second, it generates retrieval vectors for the information to be retrieved across multiple dimensions. For each dimension's retrieval vector, a search is performed in the corresponding pre-defined knowledge database, obtaining knowledge retrieval results for each dimension. These multi-dimensional knowledge retrieval results are then integrated to obtain the second corpus data. Finally, the first and second corpora are merged to obtain the enhanced corpus, which serves as semantic data for large-scale model analysis. The mechanism for identifying the authenticity of network search results based on both text-image consistency and semantic authenticity, as well as the scheme combining multi-dimensional structured knowledge retrieval enhancement generation provided in this application, can obtain more comprehensive and credible semantic data. When large models use the semantic data provided by this application to perform analysis tasks, they will inevitably obtain more credible analysis results, thereby ensuring that large models have traceable, verifiable, and anti-illusion credibility enhancement capabilities. Attached Figure Description

[0053] Figure 1 This is a flowchart of a large model event analysis enhancement method based on multi-dimensional retrieval enhancement generation, as described in one embodiment of this application;

[0054] Figure 2 This is an example of the information processing logic framework for a large model described in one embodiment of this application;

[0055] Figure 3 This is a flowchart illustrating how a network search is performed on the information to be retrieved and how the first corpus data is obtained, according to one embodiment of this application.

[0056] Figure 4 This is a flowchart illustrating, in one embodiment of the present application, the multi-dimensional search enhancement generation of the information to be retrieved and the knowledge retrieval results corresponding to each dimension;

[0057] Figure 5 This is a schematic diagram of the hardware connections of an electronic device that performs a large model event analysis enhancement method based on multi-dimensional retrieval enhancement generation, as described in one embodiment of this application. Detailed Implementation

[0058] The specific embodiments of this application will be further described below with reference to the accompanying drawings.

[0059] It is readily understood that, based on the technical solution of this application, various structural and implementation methods can be interchanged by those skilled in the art without altering the essential spirit of this application. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this application and should not be considered as the entirety of this application or as limitations or restrictions on the technical solution of the application.

[0060] This embodiment provides a method for enhancing event analysis of large models based on multi-dimensional retrieval enhancement, such as... Figure 1 As shown, the method includes the following steps:

[0061] S100: Obtain the information to be retrieved.

[0062] In this step, the information to be retrieved includes the analysis instruction prompts entered by the user and the event content to be analyzed.

[0063] S200: Based on the information to be retrieved, several online search results are obtained online; each online search result includes text and image; for each online search result, the text-image consistency evaluation result and the semantic authenticity evaluation result of the text are calculated; the first corpus data is obtained by integrating each online search result and its corresponding text-image consistency evaluation result and semantic authenticity evaluation result.

[0064] This step was performed via an online search.

[0065] S300: The information to be retrieved is input into a multi-dimensional pre-embedded model to obtain a multi-dimensional retrieval vector. The multi-dimensional vector includes at least two of the following dimensions: subject feature dimension, recent event dimension, historical event dimension, and subjective argument dimension. The retrieval vector of each dimension is used to query the corresponding dimension's pre-defined knowledge database to obtain the knowledge retrieval result corresponding to each dimension. The knowledge retrieval results of multiple dimensions are integrated to obtain the second corpus data.

[0066] In this step, the above dimensional features are provided as examples and can be adjusted according to user needs in actual applications.

[0067] S400: The first corpus data and the second corpus data are merged to obtain the enhanced corpus, which is used as semantic data for analysis by the large model.

[0068] Combination Figure 2The large-scale model information analysis logic, the process of inputting the information to be retrieved into the input interface of the large-scale model, and the process of the large-scale model text parser using the semantic data to perform semantic analysis after the enhanced corpus is used as semantic data are the same as the operations in existing large-scale models, and will not be described in detail in this application. In this application, the main improvement points of this application are described in detail as follows: after obtaining the information to be retrieved, the authenticity of the network search results for the information to be retrieved is verified, and the retrieval enhancement generation is performed from multiple dimensions.

[0069] The above-described scheme of this application involves the large model performing two-stage retrieval processes after acquiring the information to be retrieved. Firstly, it queries online search results and evaluates the consistency between the text and images and the semantic authenticity of these results, obtaining text-image consistency evaluation results and semantic authenticity evaluation results. These three evaluation results are then integrated to obtain the first corpus data. Secondly, it generates retrieval vectors for the information to be retrieved across multiple dimensions. For each dimension's retrieval vector, it searches a pre-defined knowledge database for that dimension, obtaining knowledge retrieval results for each dimension. These results from multiple dimensions are then integrated to obtain the second corpus data. Finally, the first and second corpora are merged to obtain an enhanced corpus, which serves as semantic data for the large model to analyze. The mechanism for identifying the authenticity of network search results based on both text-image consistency and semantic authenticity, as well as the scheme combining multi-dimensional structured knowledge retrieval enhancement generation provided in this application, can obtain more comprehensive and credible semantic data. When large models use the semantic data provided by this application to perform analysis tasks, they will inevitably obtain more credible analysis results, thereby ensuring that large models have traceable, verifiable, and anti-illusion credibility enhancement capabilities.

[0070] Preferably, such as Figure 3 As shown, in the above-mentioned large model event analysis enhancement method based on multi-dimensional retrieval enhancement, step S200 specifically includes:

[0071] S201: Extract keywords from the information to be retrieved, and search the internet to obtain several online search results related to the keywords.

[0072] In this step, the large model calls the standard web search API (Bing Web Search API) to retrieve web resources based on the keywords in the information to be retrieved, and obtains several search results.

[0073] S202: For each web search result, extract the text and images; wherein the text includes the title, body, source, and publication time; the images include accompanying images and video frame images, prioritizing the first image or video keyframes, in JPEG / PNG format, with the resolution uniformly scaled to 512×512. In some solutions, metadata such as URL, domain authority rating, and publishing organization type can also be extracted from the web search results.

[0074] S203: For each web search result, the text is processed into standardized text, the image is processed into standardized image, and the standardized text and the standardized image are merged to obtain a standardized image-text combination.

[0075] In this step, the text is cleaned and standardized: HTML tags, advertising fragments, and duplicate paragraphs are removed, and the text is truncated to a maximum length of 512 tokens; the image is normalized (pixel values ​​are scaled to [0,1], and the mean and variance are standardized); if a web search result does not have a valid image, the image field can be set to empty during processing.

[0076] S204: For each web search result, the standardized image-text combination and the standardized text are input in parallel into the trained image-text consistency authentication model and the semantic topic authenticity authentication model, respectively. The image-text consistency authentication model outputs the image-text consistency evaluation result, and the semantic topic authenticity authentication model outputs the semantic authenticity evaluation result. If the image field is set to empty, the image-text consistency evaluation result can be empty.

[0077] S205: Arrange each of the aforementioned network search results and its corresponding image-text consistency evaluation results and semantic authenticity evaluation results in sequence to obtain the first corpus data. Integrate the image-text consistency evaluation results and semantic authenticity evaluation results into a structured credibility description language, and append it to the end of the original network search result text in the form of invisible meta-annotations to form the first corpus data. For example, the format of the first corpus data can be:

[0078] [Online search results]<!-- TRUST_METADATA -->

[0079] [Reliability Assessment]

[0080] - Image-text consistency assessment result: C mm =0.62 (>0.55 is considered strong consistency)

[0081] - Semantic authenticity assessment results: p text =0.58 (<0.6 is considered high risk)

[0082] [Note] The above assessment results are for reference only.

[0083] <!-- / TRUST_METADATA-->

[0084] In the above format, meta-annotations are wrapped in HTML annotation syntax to ensure that they do not affect the front-end display of the large model, but can be recognized by the large model's text parser and used for decision generation. All first corpus data annotated with credibility assessment are uniformly stored in a temporary information pool, which can also be used as input for subsequent multi-dimensional retrieval enhancement generation. The large model can better understand the credibility score and can explicitly reference or implicitly perceive the above credibility assessment results when generating information analysis.

[0085] In the above-mentioned scheme of this application, the large model calls network information resources through a standardized interface and performs structured preprocessing on the obtained network search results. Then, it performs image-text consistency authentication and semantic topic authenticity authentication in parallel. Finally, the evaluation scores obtained from the two authentications are embedded into the original network search results in the form of structured prompt words to form enhanced first corpus data for subsequent large model inference.

[0086] Furthermore, in the large model event analysis enhancement method based on multi-dimensional retrieval enhancement, in step S204, the image-text consistency authentication model is obtained in the following way:

[0087] S2041: The CLIP (Contrastive Language-Image Pretraining) architecture is selected as the first initial model. The core of the CLIP architecture lies in achieving cross-modal semantic alignment between text and images through contrastive learning. The CLIP architecture's cross-modal mechanism adapts to information scenarios (such as a news article typically including an image, and the image maintaining semantic consistency with the news article). Its dual-encoder design avoids the computational overhead of an additional fusion module, while the cross-modal attention layer ensures approximate image and text encoding, thereby retrieving image-text consistency through vector similarity. Specifically, the CLIP architecture includes a text encoder, an image encoder, and a cross-modal fusion unit; wherein:

[0088] The text encoder is implemented using a BERT-based Transformer encoder (using only the encoder portion). The text encoder is used to encode the normalized text into a q-dimensional semantic vector (which can be 512). , This represents the text encoding function in the CLIP architecture, where T represents the standardized text. Let R represent the q-dimensional embedding space, and R represent the spatial domain. In specific implementation, the results of the text encoder are aggregated and output through [EOS] tokens, ensuring consistency with the image embedding space.

[0089] The image encoder is implemented based on a ResNet-50 convolutional neural network encoder. The image encoder is used to encode the normalized image (512×512 RGB image) into a q-dimensional visual vector. , This represents the image encoding function in the CLIP architecture, where I represents the normalized image. In its specific implementation, the image encoder's result is aggregated and output using the [CLS] token, ensuring consistency with the text embedding space.

[0090] The cross-modal fusion engine incorporates a task-customized cross-modal attention layer. It is used to weightedly fuse semantic vector features and visual vector features and output a fused feature vector. , h is the cross-attention operation function. T For text query features, h I Image key-value features; fused feature vectors It can be used to detect semantic conflicts. The above q acts as a scaling factor, which can prevent the dot product from becoming too large.

[0091] S2042: A first classification layer is set after the first initial model. The first classification layer is pre-trained using text-image pairs and consistency probability labels of each text-image pair from a specific dataset. Training is completed when the first difference between the consistency probability prediction result calculated by the first classification layer and the consistency probability label satisfies the convergence condition, wherein: C mm For the consistency probability prediction result and C mm ∈[0,1], It is the sigmoid activation function. ∈R 1×q For trainable weights, ∈R represents a bias term;

[0092] The first difference is calculated using the cross-entropy loss function: ; Let be the consistency probability label for the i-th text-image pair, with a value of 1 indicating consistency and a value of 0 indicating inconsistency; N is the total number of text-image pairs. This represents the consistency probability prediction result for the i-th text-image pair.

[0093] The specific dataset can be selected from commonly used datasets in existing big data analysis. In this solution, the Weibo dataset (i.e., the public dataset provided by Weibo) is chosen. In this dataset, authentic data comes from authoritative media, while inaccurate data is scraped from Weibo and refuted by official debunking systems. In specific implementation, the samples in the Weibo dataset are processed to include text-image pairs and consistency labels. Data augmentation is performed on the samples, expanding the data to approximately 8,000 or more through image cropping and text synonym replacement, thereby improving model robustness and expanding application scenarios.

[0094] S2043: The image-text consistency authentication model is obtained by encapsulating the first initial model and the first classification layer; wherein, after receiving the fused feature vector, the consistency probability prediction result output by the first classification layer is used as the image-text consistency evaluation result.

[0095] In this scheme, the image-text consistency authentication model outputs the consistency probability prediction result C. mm As a result of the text-image consistency assessment (C) mm A value >0.55 indicates strong consistency, otherwise weak consistency. Then, structured annotations are generated: [Image-Text Consistency]: C mm =0.62.

[0096] Furthermore, the semantic topic authenticity detection model in S204 of the above method is obtained in the following way:

[0097] S2044: The DeBERTa-v3 (Decomposed Embeddings with BidirectionalTransformers v3) architecture was selected as the second initial model. The DeBERTa-v3 architecture improves the ability to capture long-distance semantic dependencies and fine-grained logical contradictions by decoupling the attention mechanism and enhancing positional encoding. It is especially suitable for recognizing implicit biases and fact distortions in text.

[0098] The DeBERTa-v3 architecture includes a text feature extractor; the text feature extractor is used to: obtain a context-aware sequence of the normalized text: Where S is the sequence length and d is the hidden layer dimension, for example, d can be 768; the global semantics of the standardized text are aggregated using a weighted pooling algorithm: This allows the model to focus on key semantic segments. This represents the perceptual sequence of the i-th standardized text. This represents the perceptual sequence of the j-th standardized text. Let w represent the weighting coefficient of the i-th standardized text, where w ∈ R. d For learning parameters;

[0099] S2045: A second classification layer is set after the second initial model. The second classification layer is pre-trained using text from a specific dataset and the real labels of each text. Training is completed when the second difference between the real prediction result calculated by the second classification layer and the real label satisfies the convergence condition, wherein: ； p text For the true prediction result and p text ∈[0,1];

[0100] The second difference is calculated using the cross-entropy loss function: , Let be the authenticity label for the i-th text, with a value of 1 indicating authenticity and a value of 0 indicating inauthenticity; M is the total number of texts. The result of the authenticity prediction for the i-th text;

[0101] S2046: The semantic topic authenticity detection model is obtained by encapsulating the second initial model and the second classification layer; wherein, after receiving the standardized text, the second classification layer calculates the authenticity prediction result of the standardized text as the semantic authenticity evaluation result.

[0102] The semantic topic authenticity detection model in this solution focuses on evaluating the authenticity of textual information in terms of semantic logical consistency, factual basis for the topic, and contextual reasonableness. It does not rely on any multimedia information and identifies false or misleading content solely at the plain text level. The model is optimized to address common patterns of falsehood in information (such as unsubstantiated assertions, logical jumps, exaggerated statements, and temporal / spatial misalignments). The model output p text As the result of semantic authenticity assessment, and generating structured annotations: [Semantic Authenticity]: p text =0.58. This annotation is not used for automatic filtering, but rather as fine-grained reliable evidence for large models to comprehensively judge and confirm weights during information analysis.

[0103] In practical applications, since the vector spaces fine-tuned based on pre-embedded models of different dimensions have different characteristics, they can provide biased retrieval information during retrieval. Therefore, this scheme independently constructs a dedicated vector library and pre-embedded model for each dimension, and fine-tunes it through comparative learning to focus the vector space of each model on the semantic features of that dimension. During the retrieval phase, retrieval paths for all dimensions are triggered simultaneously, providing the large model with richer and more comprehensive semantic information. Therefore, further preferredly, such as... Figure 4 As shown, in the above scheme, the pre-set embedding model for each dimension in S300 is obtained in the following way:

[0104] S301: Select the corpus corresponding to the current dimension, and extract triples (anchor point, positive sample, negative sample) that conform to the semantics of this dimension from the corpus. Wherein:

[0105] Anchor: Describes the behavior of a subject in a specific event, such as subject A criticizing unit A's behavior problem I in a specific event;

[0106] Positive sample: Other behavioral descriptions of the same subject, such as subject A criticizing the behavior of unit / individual / group B in other events (problem II);

[0107] Negative samples: descriptions of the behavior of different subjects, such as subject B frequently engaging in activity III.

[0108] S302: Select the S-BERT architecture as the initial embedding model, wherein the S-BERT architecture includes a triplet marginal loss function; train the initial embedding model using the triplet (anchor, positive sample, negative sample), and complete the training when the triplet marginal loss result calculated by the initial embedding model satisfies the convergence condition, wherein the triplet marginal loss function is:

[0109] ;

[0110] in, This represents the anchor vector of the i-th triplet; This represents the positive sample vector of the i-th triplet. Let K represent the negative sample vector of the i-th triplet; K represents the total number of sample vectors. Indicates the marginal threshold; This represents the Euclidean distance regularization term; This represents the regularization weight.

[0111] This step uses a triplet marginal loss function to optimize the embedding space. By optimizing the triplet marginal loss function, it ensures that the same subject has similar semantics, while different subjects have different semantics. In the vector space, Euclidean distance directly reflects semantic differences (cosine similarity is affected by vector length). The marginal threshold μ has a clear physical meaning in Euclidean distance: the minimum distance difference. Euclidean distance regularization term. To prevent overfitting, regularize the weights. Balancing loss and generalization ability.

[0112] S303: The initial embedding model that has been trained is used as a preset embedding model; the preset embedding model for each dimension is used to convert the information to be retrieved into a retrieval vector for that dimension, and also to convert the knowledge information base for that dimension into a vector base.

[0113] This solution can generate multiple pre-built embedding models of different dimensions, serving as the basis for multi-dimensional retrieval enhancement generation.

[0114] Therefore, after the information to be retrieved is input into pre-defined embedding models of different dimensions, each pre-defined embedding model of a given dimension can produce a retrieval vector biased towards that dimension. Taking a pre-defined embedding model of five dimensions (subjective features, recent events, rule basis, historical events, and subjective arguments) as an example, if the information to be retrieved Q is simultaneously input into the pre-defined embedding models of all five dimensions, each pre-defined embedding model of a given dimension independently encodes Q into a retrieval vector q for that dimension. d , d∈{1,2,3,4,5}.

[0115] In S300 of the above scheme: the pre-set knowledge database for each dimension includes a knowledge information database, a vector database, and a mapping relationship for that dimension, wherein: the knowledge information database includes knowledge information for that dimension; the pre-set embedding model processes each piece of knowledge information to obtain a knowledge vector, and the knowledge vectors form the vector database; the mapping relationship is the mapping relationship between each knowledge vector in the vector database and the corresponding knowledge information in the knowledge information database. That is, in practical applications, the pre-set embedding model processes the knowledge information database to obtain the vector database, and constructs a vector index, which then indexes all vectors in the vector database. Construct an approximate nearest neighbor index. Specifically, the HNSW (Hierarchical Navigable Small World) algorithm can be used to construct the index, with the index type being hnsw:space=cosine,ef_construction=200,M=16, and the storage format supporting high recall and low latency being set to FAISS.

[0116] Furthermore, such as Figure 4 As shown, in step S300, the retrieval vector for each dimension is used to query the pre-set knowledge database corresponding to that dimension to obtain the knowledge retrieval result for each dimension. The knowledge retrieval results from multiple dimensions are then integrated to obtain the second corpus data, including:

[0117] S304: Query the vector library of the preset knowledge database for the knowledge vector that has the highest similarity to the retrieved vector as the target vector.

[0118] S305: Based on the mapping relationship, search the knowledge information in the knowledge information base that corresponds to the target vector as the target knowledge information.

[0119] S306: The knowledge retrieval result for that dimension is obtained by combining the dimension and the target knowledge information.

[0120] Based on the foregoing, the retrieval vector q for each dimension d A nearest neighbor search is performed in the vector library for the corresponding dimension, and the closest knowledge vector found is indexed into the original knowledge information. In practical applications, a specific number can be limited to reduce the amount of information processing; for example, three of the closest target knowledge information are selected for each dimension, which can be adjusted according to the actual situation. If there are five dimensions, a total of 15 target knowledge information will be selected.

[0121] In practical applications, the retrieved target knowledge information can be further processed in three ways to ensure information quality and dimensionality isolation:

[0122] Semantic segmentation: The knowledge information in the knowledge information base is segmented into independent blocks of ≤256 tokens according to semantic boundaries, and the source metadata (title, publishing organization, authority rating, timestamp) is preserved.

[0123] Cross-dimensional deduplication: If the same knowledge information is retrieved in multiple dimensions (such as the historical event dimension and the rule basis dimension), only its first appearance in the dimension with the highest similarity (such as the historical event dimension similarity 0.83 > the rule basis dimension 0.78) is retained, and the other dimensions are marked as "covered".

[0124] Confidence filtering: Based on the similarity calculation in S304, results with similarity <0.80 are filtered to ensure that only query results with high confidence are retained.

[0125] In step S306, the second corpus obtained can be achieved through direct concatenation, resulting in the following format:

[0126] [MULTI-DIMENSIONAL EVIDENCE]

[0127] [Subject Feature Dimension]

[0128] Person A recently spoke out regarding Company A's controversial behavior I;

[0129] Person A is deeply involved in the field of monitoring controversial events, with a core characteristic being a focus on industry compliance and safety issues in areas a, b, and c.

[0130] Person A has a certain public influence and is good at speaking out through public social channels to attract attention;

[0131] Company B is a mid-range mass-market chain brand, mainly engaged in Behavior II, Behavior III, and Behavior IV, with offline stores covering multiple cities;

[0132] [Recent Events]

[0133] Recently, Person A published multiple posts on public social media channels regarding controversial behavior I of Company A. These posts included: Post A1, Post A2, ... Post An; the content quickly sparked widespread discussion and attention.

[0134] The day after the content was published, Company A issued a public response and denied the allegations, which included: Content A1, Content A2... Content Am; Company A and Person A had a disagreement on the controversial issues;

[0135] Subsequently, Company A continued to issue public responses, including: Content A11, Content A12... Content A1k; Company A and Person A continue to have disagreements on the controversial issues;

[0136] [Historical Events Dimension]

[0137] During its past operations, Company A has experienced several incidents related to controversial behavior I, such as: incidents Ia, Ib...Im;

[0138] Company A's industry has seen numerous public reports in recent years regarding controversial behaviors, such as events Il, c, Ild...Iln;

[0139] [Subjective Argumentation Dimension]

[0140] Industry experts from Company A commented: The core of this disagreement lies in H1, H2...Hm, and their suggestions include J1, J2...Jn;

[0141] Feedback from ordinary consumers: The controversial event itself is the core concern of consumers, and the expected outcome includes W1, W2...Wk.

[0142] <!-- / MULTI-DIMENSIONAL EVIDENCE -->

[0143] After multi-dimensional retrieval is completed, step S400 merges the first and second corpus data to obtain enhanced corpus. This involves deep, structured fusion of the first and second corpus data, seamlessly integrating them into a unified, machine-readable input, ensuring that the large model can directly obtain semantically coherent and reliable enhanced corpus. The specific fusion process is as follows:

[0144] The processed second corpus data is integrated into a machine-parseable structured data package according to dimensional priority (e.g., main features → recent events → historical events → subjective arguments). The format follows the four-tuple data package of "dimensional label-knowledge information content-confidence-source". This structured data package is placed after the first corpus data to form the final enhanced corpus, with the following format:

[0145] [Original search information: Person A's evaluation of Company A: Controversial behavior I!]

[0146] <!-- TRUST_METADATA -->

[0147] [Reliability Assessment]

[0148] Image-text consistency assessment result: C mm =0.62 (>0.55 is considered strong consistency);

[0149] Semantic authenticity assessment results: p text =0.58 (<0.6 is considered high risk);

[0150] [Note] The above assessment results are for reference only;

[0151] <!-- / TRUST_METADATA-->

[0152] [MULTI-DIMENSIONAL EVIDENCE]

[0153] [Subject Feature Dimension]

[0154] Person A recently spoke out regarding Company A's controversial behavior I;

[0155] Person A is deeply involved in the field of monitoring controversial events, with a core characteristic being a focus on industry compliance and safety issues in areas a, b, and c.

[0156] Person A has a certain public influence and is good at speaking out through public social channels to attract attention;

[0157] Company B is a mid-range mass-market chain brand, mainly engaged in Behavior II, Behavior III, and Behavior IV, with offline stores covering multiple cities;

[0158] [Recent Events]

[0159] Recently, Person A published multiple posts on public social media channels regarding controversial behavior I of Company A. These posts included: Post A1, Post A2, ... Post An; the content quickly sparked widespread discussion and attention.

[0160] The day after the content was published, Company A issued a public response and denied the allegations, which included: Content A1, Content A2... Content Am; Company A and Person A had a disagreement on the controversial issues;

[0161] Subsequently, Company A continued to issue public responses, including: Content A11, Content A12... Content A1k; Company A and Person A continue to have disagreements on the controversial issues;

[0162] [Historical Events Dimension]

[0163] During its past operations, Company A has experienced several incidents related to controversial behavior I, such as: incidents Ia, Ib...Im;

[0164] Company A's industry has seen numerous public reports in recent years regarding controversial behaviors, such as events Il, c, Ild...Iln;

[0165] [Subjective Argumentation Dimension]

[0166] Industry experts from Company A commented: The core of this disagreement lies in H1, H2...Hm, and their suggestions include J1, J2...Jn;

[0167] Feedback from ordinary consumers: The controversial event itself is the core concern of consumers, and the expected outcome includes W1, W2...Wk.

[0168] <!-- / MULTI-DIMENSIONAL EVIDENCE -->

[0169] The above is an example format for the final enhanced corpus.

[0170] The large-model event analysis enhancement method based on multi-dimensional retrieval enhancement provided in this application achieves the following core advantages through the deep integration of a dual-model collaborative authentication mechanism and multi-dimensional retrieval enhancement generation:

[0171] (1) Enhanced authentication module: The dual-model collaborative evaluation of the image and text consistency authentication model and the semantic topic authenticity authentication model is adopted. The credibility evaluation results are embedded into the network search results with structured annotations to avoid the general search enhancement generation from fully accepting or partially filtering false search content.

[0172] (2) Breakthrough in multidimensional retrieval enhancement generation: predefine multiple orthogonal analysis dimensions, construct a dedicated embedding model for each dimension for parallel retrieval, solve the analysis bias caused by the single dimension of traditional retrieval enhancement generation, and ensure that the large model synchronously acquires multi-source information such as behavioral logic, historical context, and basis.

[0173] The semantic analysis of the enhanced corpus described above can be achieved using existing methods. However, since the enhanced corpus provided in this proposal has a credibility assessment result, and it also proposes multi-dimensional retrieval enhancement generation, the semantic analysis of the enhanced corpus based on this application will inevitably yield more credible results.

[0174] This application provides a computer-readable storage medium storing program information. After reading the program information, the computer executes the steps of the large model event analysis enhancement method based on multi-dimensional retrieval enhancement generation described in any of the above method embodiments.

[0175] This application provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the large model event analysis enhancement method based on multi-dimensional retrieval enhancement generation described in any of the above method embodiments.

[0176] This application also provides an electronic device, such as... Figure 5 As shown, the electronic device includes at least one processor 501 and at least one memory 502. The at least one memory 502 stores program information. After reading the program information, the at least one processor 501 executes the large model event analysis enhancement method based on multi-dimensional retrieval enhancement generation as described in any of the above method embodiments. The device may further include an input device 503 and an output device 504. The processor 501, memory 502, input device 503, and output device 504 can be communicatively connected. The memory 502, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules. The processor 501 executes various functional applications and data processing by running the non-volatile software programs, instructions, and modules stored in the memory 502, thereby implementing the large model event analysis enhancement method based on multi-dimensional retrieval enhancement generation provided in any of the above embodiments. The memory 502 may include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a function; the data storage area may store data created based on the use of the large model event analysis enhancement method based on multi-dimensional retrieval enhancement generation. Furthermore, memory 502 may include high-speed random access memory and non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some embodiments, memory 502 may optionally include memory remotely located relative to processor 501, and these remote memories may be connected via a network to a device performing a large model event analysis enhancement method based on multi-dimensional retrieval enhancement generation. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof. Input device 503 may receive user clicks and generate signal inputs related to user settings and function control of the large model event analysis enhancement method based on multi-dimensional retrieval enhancement generation. Output device 504 may include a display device such as a display screen. When the one or more modules are stored in memory 502 and are run by the one or more processors 501, the large model event analysis enhancement method based on multi-dimensional retrieval enhancement generation described in any of the above method embodiments is executed.

[0177] As needed, the above technical solutions can be combined to achieve the best technical effect.

[0178] The above are merely the principles and preferred embodiments of this application. It should be noted that, for those skilled in the art, several other modifications can be made based on the principles of this application, and these modifications should also be considered within the scope of protection of this application.

Claims

1. A method for enhancing event analysis in large models based on multi-dimensional retrieval enhancement, characterized in that, include: S100: Obtain the information to be retrieved; S200: Based on the information to be retrieved, several online search results are obtained online; Each web search result includes text and image; for each web search result, the text-image consistency evaluation result and the semantic authenticity evaluation result of the text are calculated; the first corpus data is obtained by integrating each web search result and its corresponding text-image consistency evaluation result and semantic authenticity evaluation result. S300: The information to be retrieved is input into a multi-dimensional preset embedding model to obtain a multi-dimensional retrieval vector. The multiple dimensions include at least two of the following: subject feature dimension, recent event dimension, historical event dimension, and subjective argument dimension. The retrieval vector of each dimension is used to query the corresponding dimension's preset knowledge database to obtain the knowledge retrieval result corresponding to each dimension. The knowledge retrieval results of multiple dimensions are integrated to obtain the second corpus data. S400: The first corpus data and the second corpus data are merged to obtain the enhanced corpus, which is used as semantic data for analysis by the large model.

2. The method for enhancing large-scale model event analysis based on multi-dimensional retrieval enhancement as described in claim 1, characterized in that, Step S200 specifically includes: S201: Extract keywords from the information to be retrieved, and search the internet to obtain several online search results related to the keywords; S202: For each web search result, extract the text and images; wherein the text includes the title, body text, source, and publication time; and the images include accompanying pictures and video frame images. S203: For each web search result, the text is processed into standardized text, the image is processed into standardized image, and the standardized text and the standardized image are merged to obtain a standardized image-text combination; S204: Input the standardized image-text combination and the standardized text corresponding to each network retrieval result into the trained image-text consistency detection model and semantic topic authenticity detection model in parallel. The image-text consistency detection model outputs the image-text consistency evaluation result, and the semantic topic authenticity detection model outputs the semantic authenticity evaluation result. S205: Arrange each of the network retrieval results and its corresponding image-text consistency evaluation results and semantic authenticity evaluation results in order to obtain the first corpus data.

3. The method for enhancing large-scale model event analysis based on multi-dimensional retrieval enhancement as described in claim 2, characterized in that, In step S204, the image-text consistency authentication model is obtained in the following way: S2041: Select the CLIP architecture as the first initial model, wherein the CLIP architecture includes a text encoder, an image encoder, and a cross-modal fusion unit; wherein: The text encoder is used to encode the standardized text into a q-dimensional semantic vector: , This represents the text encoding function, where T represents the standardized text. Let R denote the q-dimensional embedding space, and let R denote the spatial domain; The image encoder is used to encode the normalized image into a q-dimensional visual vector: , I represents the image coding function, and I represents the normalized image. The cross-modal fusion engine is used to weight and fuse semantic vector features and visual vector features and output a fused feature vector: , h is the cross-attention operation function. T For text query features, h I Image key-value features; S2042: A first classification layer is set after the first initial model. The first classification layer is pre-trained using text-image pairs and consistency probability labels of each text-image pair from a specific dataset. Training is completed when the first difference between the consistency probability prediction result calculated by the first classification layer and the consistency probability label satisfies the convergence condition, wherein: C mm For the consistency probability prediction result and C mm ∈[0,1], It is the sigmoid activation function. ∈R 1×q For trainable weights, ∈R represents a bias term; The first difference is calculated using the cross-entropy loss function: ; Let be the consistency probability label for the i-th text-image pair, with a value of 1 indicating consistency and a value of 0 indicating inconsistency; N is the total number of text-image pairs. The consistency probability prediction result for the i-th text-image pair; S2043: The image-text consistency authentication model is obtained by encapsulating the first initial model and the first classification layer; wherein, after receiving the fused feature vector, the consistency probability prediction result output by the first classification layer is used as the image-text consistency evaluation result.

4. The method for enhancing large-scale model event analysis based on multi-dimensional retrieval enhancement as described in claim 3, characterized in that, In step S204, the semantic topic authenticity detection model is obtained in the following way: S2044: Select the DeBERTa-v3 architecture as the second initial model, which includes a text feature extractor; the text feature extractor is used to: obtain a context-aware sequence of the normalized text: Where S is the sequence length and d is the hidden layer dimension; the global semantics of the standardized text are aggregated using a weighted pooling algorithm: ;in, This represents the perceptual sequence of the i-th standardized text. This represents the perceptual sequence of the j-th standardized text. Let w represent the weighting coefficient of the i-th standardized text, where w ∈ R. d For learning parameters; S2045: A second classification layer is set after the second initial model. The second classification layer is pre-trained using text from a specific dataset and the real labels of each text. Training is completed when the second difference between the real prediction result calculated by the second classification layer and the real label satisfies the convergence condition, wherein: ； p text For the true prediction result and p text ∈[0,1]; The second difference is calculated using the cross-entropy loss function: , Let be the authenticity label for the i-th text, with a value of 1 indicating authenticity and a value of 0 indicating inauthenticity; M is the total number of texts. The result of the authenticity prediction for the i-th text; S2046: The semantic topic authenticity detection model is obtained by encapsulating the second initial model and the second classification layer; wherein, after receiving the standardized text, the second classification layer outputs the authenticity prediction result as the semantic authenticity evaluation result.

5. The method for enhancing large-scale model event analysis based on multi-dimensional retrieval enhancement according to any one of claims 1-4, characterized in that, In S300, the pre-set embedding model for each dimension is obtained in the following way: S301: Select a corpus corresponding to the current dimension, and extract triples that conform to the semantics of the dimension from the corpus. The triples include anchor points, positive samples, and negative samples. S302: Select the S-BERT architecture as the initial embedding model, wherein the S-BERT architecture includes a triplet marginal loss function; train the initial embedding model using the triplets, and complete the training when the triplet marginal loss result calculated by the initial embedding model satisfies the convergence condition, wherein the triplet marginal loss function is: ； in, This represents the anchor vector of the i-th triplet; This represents the positive sample vector of the i-th triplet. Let K represent the negative sample vector of the i-th triplet; K represents the total number of sample vectors. Indicates the marginal threshold; This represents the Euclidean distance regularization term; Indicates the regularization weight; S303: Use the initial embedding model that has been trained as a preset embedding model; The pre-set embedding model for each dimension is used to convert the information to be retrieved into a retrieval vector for that dimension, and also to convert the knowledge information base for that dimension into a vector base.

6. The method for enhancing large-scale model event analysis based on multi-dimensional retrieval enhancement as described in claim 5, characterized in that, In S300: The pre-built knowledge database for each dimension includes a knowledge information database, a vector database, and mapping relationships for that dimension, wherein: The knowledge information base includes knowledge information under this dimension; Each piece of knowledge information is processed using the pre-set embedding model to obtain a knowledge vector, and the knowledge vectors form the vector library. The mapping relationship is the mapping relationship between each knowledge vector in the vector library and the corresponding knowledge information in the knowledge information library.

7. The method for enhancing large-scale model event analysis based on multi-dimensional retrieval enhancement as described in claim 6, characterized in that, In step S300, the retrieval vector for each dimension is used to query the pre-set knowledge database corresponding to that dimension to obtain the knowledge retrieval result for each dimension. The knowledge retrieval results from multiple dimensions are then integrated to obtain the second corpus data, including: S304: Query the vector library of the preset knowledge database for the knowledge vector that has the highest similarity to the retrieval vector as the target vector; S305: Based on the mapping relationship, search the knowledge information in the knowledge information base that corresponds to the target vector as the target knowledge information; S306: The knowledge retrieval result for that dimension is obtained by combining the dimension and the target knowledge information.

8. A computer-readable storage medium, characterized in that, The storage medium stores program information, and after the computer reads the program information, it executes the steps of the large model event analysis enhancement method based on multi-dimensional retrieval enhancement as described in any one of claims 1-7.

9. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the steps of the large model event analysis enhancement method based on multi-dimensional retrieval enhancement generation as described in any one of claims 1-7.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the large model event analysis enhancement method based on multi-dimensional retrieval enhancement generation as described in any one of claims 1-7.

Images