CN-121996986-A - Large model event analysis enhancement method based on multidimensional retrieval enhancement generation

CN121996986ACN 121996986 ACN121996986 ACN 121996986ACN-121996986-A

Abstract

The application provides a large model event analysis enhancement method based on multi-dimensional retrieval enhancement generation, and relates to the technical field of artificial intelligence and information analysis intersection. After the information to be searched is obtained, on one hand, a network search result is inquired on the network, an image-text consistency evaluation result and a semantic authenticity evaluation result are obtained according to the network search result, and the network search result, the image-text consistency evaluation result and the semantic authenticity evaluation result are integrated to obtain first corpus data. On the other hand, the information to be searched generates search vectors in a plurality of dimensions, knowledge search results of each dimension are obtained by searching in a corresponding preset knowledge database, and the knowledge search results of the plurality of dimensions are integrated to obtain second corpus data. And combining the first corpus data and the second corpus data to obtain enhanced corpus which is used as semantic material for analysis of the large model. When the large model uses the semantic data provided by the scheme of the application to execute the analysis task, the analysis result with higher credibility can be obtained.

Inventors

LI XIANNENG
WANG JUE
SONG ZILONG
PENG HUANXIN
YANG SIQI
LI ZHIPENG
WANG YITAO
HU DEQIANG
YU YANG
ZHANG ZHONGZHAO

Assignees

大连理工大学

Dates

Publication Date: 20260508
Application Date: 20260403

Claims (10)

1. The large model event analysis enhancement method based on multi-dimensional retrieval enhancement generation is characterized by comprising the following steps of: S100, obtaining information to be retrieved; S200, inquiring on a network according to the information to be searched to obtain a plurality of network search results, wherein each network search result comprises a text and an image, and calculating an image-text consistency evaluation result of the text and the image and a semantic authenticity evaluation result of the text according to each network search result; S300, respectively inputting the information to be searched into a preset embedding model with multiple dimensions to obtain search vectors with multiple dimensions, wherein the multiple dimensions comprise at least two of a main feature dimension, a recent event dimension, a historical event dimension and a subjective discussion dimension; s400, merging the first corpus data and the second corpus data to obtain enhanced corpus, wherein the enhanced corpus is used as semantic material for analysis of a large model.
2. The method for enhancing analysis of large model events based on multi-dimensional search enhancement generation according to claim 1, wherein step S200 specifically comprises: S201, extracting keywords in the information to be searched, and inquiring on a network to obtain a plurality of network search results related to the keywords; S202, extracting text and images from each network search result, wherein the text comprises a title, a body, a source and release time; S203, processing the text into a standardized text, processing the image into a standardized image according to each network search result, and merging the standardized text and the standardized image to obtain a standardized image-text combination; S204, inputting the standardized image-text combination and the standardized text corresponding to each network search result into an image-text consistency authentication model and a semantic topic authenticity authentication model which are subjected to training in parallel, wherein 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; S205, sequentially arranging each network search result, the corresponding image-text consistency evaluation result and the semantic authenticity evaluation result to obtain the first corpus data.
3. The method for analyzing and enhancing a large model event generated based on multi-dimensional search enhancement according to claim 2, wherein in S204, the graph-text consistency authentication model is obtained by: s2041, selecting a CLIP architecture as a first initial model, wherein the CLIP architecture comprises a text encoder, an image encoder and a cross-modal fusion device, and the CLIP architecture comprises: the text encoder is configured to encode the normalized text into a q-dimensional semantic vector: , Representing a text encoding function, T representing the normalized text, Representing q-dimensional embedding space, R representing a spatial domain; The image encoder is configured to encode the normalized image into a q-dimensional visual vector: , representing an image coding function, I representing the normalized image; The cross-modal fusion device is used for weighting and fusing semantic vector features and visual vector features and outputting fusion feature vectors: , H T is a text query feature, and h I is an image key value feature; S2042, setting a first classification layer behind the first initial model, wherein the first classification layer adopts text-image pairs in a specific data set and consistency probability labels of the text-image pairs to train in advance, and training is completed when a first difference between a consistency probability prediction result calculated by the first classification layer and the consistency probability labels meets a convergence condition, wherein: C mm is the result of the consistency probability prediction and C mm E [0,1], For the sigmoid activation function, E R 1×q is a trainable weight, E, R is a bias term; the first difference is calculated by a cross entropy loss function: ; a consistency probability label for the ith text-image pair, wherein a value of 1 indicates consistency and a value of 0 indicates non-consistency; a consistency probability predictor for an ith text-image pair; S2043, packaging the first initial model and the first classification layer to obtain the image-text consistency authentication model, wherein the first classification layer receives the fusion feature vector and then outputs a consistency probability prediction result as the image-text consistency evaluation result.
4. The method for enhancing analysis of large model events generated based on multi-dimensional search enhancement according to claim 3, wherein in S204, the semantic topic authenticity discrimination model is obtained by: s2044 selecting DeBERTa-v3 architecture as a second initial model, comprising a text feature extractor for obtaining a context aware sequence of the normalized text: wherein S is the sequence length, d is the hidden layer dimension, and the global semantics of the standardized text are aggregated through a weighted pooling algorithm: wherein, the method comprises the steps of, Representing the perceived sequence of the ith normalized text, Representing the perceived sequence of the jth normalized text, The weighting coefficient of the ith standardized text is represented, and w epsilon R d is a learning parameter; S2045, setting a second classification layer behind the second initial model, wherein the second classification layer adopts texts in a specific data set and authenticity labels of the texts to train in advance, and training is completed when a second difference between an authenticity prediction result calculated by the second classification layer and the authenticity labels meets a convergence condition, wherein: P text is the result of the authenticity prediction and p text ε [0,1]; The second difference is calculated by a cross entropy loss function: , An authenticity label for the ith text, which takes a value of 1 to indicate authenticity and 0 to indicate unreliability, M is the total number of texts, The result is the authenticity prediction result of the ith text; S2046, packaging the second initial model and the second classification layer to obtain the semantic topic authenticity identification model, wherein the second classification layer receives the standardized text and outputs an authenticity prediction result as the semantic authenticity assessment result.
5. The method for large model event analysis and enhancement based on multi-dimensional search enhancement generation according to any one of claims 1 to 4, wherein in S300, the preset embedded model of each dimension is obtained by: S301, selecting a corpus corresponding to a current dimension, and extracting a triplet conforming to the dimension semantics from the corpus, wherein the triplet comprises an anchor point, a positive sample and a negative sample; S302, selecting an S-BERT framework as an initial embedding model, wherein the S-BERT framework comprises a triplet marginal loss function, training the initial embedding model by utilizing the triplet, and completing training when a triplet marginal loss result calculated by the initial embedding model meets a convergence condition, wherein the triplet marginal loss function is as follows: ; Wherein, the An anchor vector representing an i-th triplet; representing the positive sample vector of the i-th triplet, K represents the total number of sample vectors; representing a marginal threshold; representing euclidean distance regularization terms; Representing regularization weights; S303, taking the initial embedded model after training as a preset embedded model; the preset embedded model of each dimension is used for converting the information to be searched into a search vector of the dimension and converting a knowledge information base of the dimension into a vector base.
6. The method for large model event analysis enhancement based on multi-dimensional search enhancement generation according to claim 5, wherein in S300: The preset knowledge database of each dimension comprises a knowledge information base, a vector base and a mapping relation under the dimension, wherein: the knowledge information base comprises knowledge information under the dimension; processing each piece of knowledge information by adopting the preset embedding model to obtain a knowledge vector, wherein the knowledge vector forms the vector library; The mapping relation is the mapping relation between each knowledge vector in the vector library and the corresponding knowledge information in the knowledge information library.
7. The method for analyzing and enhancing a large model event generated based on multi-dimensional search enhancement according to claim 6, wherein in S300, the searching in a preset knowledge database of each dimension by using the search vector of each dimension to obtain a knowledge search result corresponding to each dimension, integrating the knowledge search results of multiple dimensions to obtain second corpus data includes: S304, inquiring a knowledge vector with the highest similarity with the search vector in a vector library of the preset knowledge database as a target vector; S305, searching knowledge information corresponding to the target vector in the knowledge information base according to the mapping relation to serve as target knowledge information; s306, combining the dimension and the target knowledge information to obtain the knowledge retrieval result of the dimension.
8. A computer-readable storage medium, wherein program information is stored in the storage medium, and a computer executes the steps of the large model event analysis enhancement method based on multi-dimensional search enhancement generation according to any one of claims 1 to 7 after reading the program information.
9. A computer program product comprising computer programs/instructions which when executed by a processor implement the steps of the large model event analysis enhancement method generated based on multi-dimensional search enhancement of any of claims 1-7.
10. An electronic device comprising a memory, a processor and a computer program stored on the memory, characterized in that the processor executes the computer program to implement the steps of the large model event analysis enhancement method generated based on multi-dimensional search enhancement of any of claims 1-7.

Description

Large model event analysis enhancement method based on multidimensional retrieval enhancement generation Technical Field The application relates to the technical field of intersection of artificial intelligence and information analysis, in particular to a large model event analysis enhancement method based on multi-dimensional retrieval enhancement generation. Background The use of large language models (Large Language Model, LLM, hereinafter referred to as large models) is widespread, relying on the ability of external retrieval enhancement generation (RETRIEVAL-Augmented Generation, RAG) to obtain semantic data is a key support. However, the standard RAG architecture faces serious reliability challenges, on one hand, the network search and retrieval sources are open networks, the content quality is uneven, misleading or unilateral information is very easy to introduce, and when the network search and retrieval sources are used as semantic data to be input into a large model, the large model can generate 'illusion' output containing fact errors or logic contradictions, so that the reliability of analysis conclusion is weakened. On the other hand, the current large model output commonly adopts a single-channel RAG framework driven by a universal embedded model, namely, a single text encoder (such as S-BERT) is used for vectorizing the whole knowledge base, and then a plurality of document fragments with the most similar retrieval semantics are used as context input large models according to user query. The method has the following defects: First, there is a lack of a mechanism for evaluating the credibility of the retrieved content. The existing RAG system generally assumes that the retrieval result is wholly trusted or wholly unreliable, namely when the retrieval result is considered to be wholly trusted, the retrieval result is directly spliced into a prompt word as semantic data for use by a large model, no fake identification processing is carried out on the original information in any form, and when the unreal information is retrieved, the large model can be used for reasoning as real evidence, so that errors are amplified. When the whole is considered to be not trusted, the RAG cannot recognize and retain the trusted fragments even if the information part is authentic (e.g. body, time correct but event details are kneaded), but directly negates the whole information. This approach results in an "all or nothing" information usage pattern, which may introduce either erroneous content or ignore potentially valid clues due to excessive caution. Secondly, the retrieval dimension is single, and complex information reasoning is difficult to support. Some types of event analysis need to integrate multidimensional heterogeneous information such as principal behavior characteristics, historical event venation, specific rule basis, expert subjective discussion and the like. However, the general embedded model can only capture the semantic similarity of the surface, and cannot distinguish the orthogonal dimensions, so that the analysis view angle is narrow and the conclusion is one-sided. Thus, a new large model information analysis trust enhancement scheme is needed to solve the above-mentioned problems. Disclosure of Invention The technical problem to be solved by the application is that the existing technical scheme for searching and enhancing generation has poor false discrimination capability and single searching dimension, which results in low reliability of large model information analysis, and further provides a large model event analysis enhancing method based on multi-dimensional searching and enhancing generation. In a first aspect, the present application provides a method for enhancing analysis of a large model event based on multi-dimensional search enhancement, including: S100, obtaining information to be retrieved; S200, inquiring on a network according to the information to be searched to obtain a plurality of network search results, wherein each network search result comprises a text and an image, and calculating an image-text consistency evaluation result of the text and the image and a semantic authenticity evaluation result of the text according to each network search result; S300, respectively inputting the information to be searched into a preset embedding model with multiple dimensions to obtain search vectors with multiple dimensions, wherein the multiple dimensions comprise at least two of a main feature dimension, a recent event dimension, a historical event dimension and a subjective discussion dimension; s400, merging the first corpus data and the second corpus data to obtain enhanced corpus, wherein the enhanced corpus is used as semantic material for analysis of a large model. Preferably, the large model event analysis enhancement method based on multi-dimensional retrieval enhancement generation, the step S200 specifically includes: S201, extracting keywords in the information to b