Object analysis method, device, computer device, readable storage medium and program product

Abstract

This application relates to an object analysis method, apparatus, computer device, computer-readable storage medium, and program product. The method includes: obtaining the original text vector corresponding to the target object; the original text vector includes the target information source corresponding to the original text vector; determining the target description vector corresponding to the target information source based on the correspondence between candidate information sources and candidate description vectors; the candidate information sources are multiple pre-stored information publishing platforms or channels, each candidate information source containing a corresponding candidate description vector, the candidate description vector being a feature supplement to each candidate information source; analyzing and processing the target description vector and the original text vector based on a sub-model of a preset large language model to obtain a first object sentiment vector of the original text vector; and determining the analysis result corresponding to the target object based on the target description vector, the original text vector, the first object sentiment vector, and the preset large language model. Using this method can improve the accuracy of the analysis results.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a data analysis method, apparatus, computer equipment, computer-readable storage medium, and computer program product. Background Technology

[0002] With the development of computer technology, various methods for analyzing the development trends of target objects based on natural language processing technology have emerged.

[0003] Traditional technologies often employ Natural Language Processing (NLP) to analyze content. For example, the target might be market trends or a patient's health status, with input including market-related news or patient test reports. A typical NLP workflow involves first segmenting the text, classifying and counting the frequency of each word, and then, based on the word classifications and the quantity of corresponding words in the text, judging audience feedback or the patient's health condition, and using existing rules to determine trends.

[0004] However, this method of analyzing text information based on traditional linguistic knowledge requires a large amount of prior knowledge and accumulation in the corresponding field, and its scope of consideration is limited, resulting in limited effectiveness and an inability to be quickly transferred to other fields.

[0005] Currently, there are also some sentiment analysis and trend analysis methods based on large language models. These methods utilize the strong classification capabilities of large language models to evaluate the input text in a human-defined direction, and then use prior knowledge to perform trend analysis and sentiment judgment based on these results. However, because the judgment is made in a human-defined direction, the high uncertainty of various input information and the incompleteness of the human-introduced directions lead to poor accuracy. Summary of the Invention

[0006] Therefore, it is necessary to provide an object analysis method, apparatus, computer device, computer-readable storage medium, and computer program product to address the aforementioned technical problems.

[0007] Firstly, this application provides an object analysis method, including:

[0008] Obtain the original text vector corresponding to the target object; the original text vector includes the target information source corresponding to the original text vector.

[0009] The target description vector corresponding to the target information source is determined according to the correspondence between the candidate information sources and the candidate description vectors; the candidate information sources are multiple information publishing platforms or channels stored in advance, and each candidate information source contains the corresponding candidate description vector, which is a feature supplement for each candidate information source;

[0010] The sub-model based on the preset large language model analyzes and processes the target description vector and the original text vector to obtain the first object sentiment vector of the original text vector;

[0011] The analysis results corresponding to the target object are determined based on the target description vector, the original text vector, the first object sentiment vector, and the preset large language model.

[0012] In one embodiment, before obtaining the original text vector corresponding to the target object, the method further includes:

[0013] Obtain the initial description vector to be trained corresponding to multiple candidate information sources, the sample data corresponding to each candidate information source, and the first sample label and the second sample label corresponding to each sample data;

[0014] Based on the preset large language model, the sample data corresponding to each candidate information source and the first sample label corresponding to each sample data, the initial description vector corresponding to each candidate information source is trained to obtain multiple intermediate candidate description vectors.

[0015] The sample data is analyzed and processed based on the trained sub-model to obtain the second object sentiment vector corresponding to each sample data.

[0016] Based on the preset large language model, the sample data, the first sample label corresponding to each sample data and the second object emotion vector, the intermediate candidate description vector corresponding to each candidate information source is retrained to obtain multiple candidate description vectors.

[0017] In one embodiment, the step of training the initial description vector corresponding to each candidate information source based on a preset large language model, sample data corresponding to each candidate information source, and first sample labels corresponding to each sample data to obtain multiple intermediate candidate description vectors includes:

[0018] For each candidate information source, the sample data corresponding to each candidate information source and the initial description vector are analyzed and processed according to the preset large language model to obtain the initial analysis results;

[0019] Based on the initial analysis results and the first sample label, determine the first loss value corresponding to the initial description vector;

[0020] The initial description vector is adjusted based on the first loss value to obtain a new initial description vector, until the first loss value corresponding to the new initial description vector satisfies the first loss condition, thus obtaining multiple intermediate candidate description vectors.

[0021] In one embodiment, before analyzing and processing the sample data based on the trained sub-model to obtain the second object emotion vector corresponding to each sample data, the method further includes:

[0022] For each sample data, the sample data is concatenated with the intermediate candidate description vector corresponding to the sample data to obtain the first input vector;

[0023] Based on the sub-models of the preset large language model, each of the first input vectors is compared and analyzed with the preset historical description text to obtain the initial emotion vector;

[0024] The sub-model is trained based on the initial emotion vector and the second sample label to obtain the trained sub-model.

[0025] In one embodiment, training the sub-model based on the initial emotion vector and the second sample label to obtain the trained sub-model includes:

[0026] The second loss value of the sub-model is determined based on the initial sentiment vector and the second sample label;

[0027] The model parameters of the sub-model are adjusted according to the second loss value until the second loss value meets the second loss condition, thus obtaining the trained sub-model.

[0028] In one embodiment, the sub-model includes historical descriptive text corresponding to the target object; the sub-model based on a preset large language model analyzes and processes the target description vector and the original text vector to obtain a first object sentiment vector of the original text vector, including:

[0029] The target description vector and the original text vector are concatenated to obtain the second input vector;

[0030] The historical descriptive text and the second input vector are compared and analyzed according to the sub-model. Based on the comparison and analysis results, the influence of the original text vector on the state of the target object is determined, and the first object emotion vector corresponding to the original text vector is obtained.

[0031] Secondly, this application also provides an object analysis apparatus, comprising:

[0032] The first acquisition module is used to acquire the original text vector corresponding to the target object; the original text vector includes the target information source corresponding to the original text vector.

[0033] The determining module is used to determine the target description vector corresponding to the target information source based on the correspondence between candidate information sources and candidate description vectors; the candidate information sources are multiple information publishing platforms or channels stored in advance, each candidate information source contains the corresponding candidate description vector, and the candidate description vector is a feature supplement for each candidate information source;

[0034] The first analysis module is used to analyze and process the target description vector and the original text vector based on the sub-model of the preset large language model to obtain the first object sentiment vector of the original text vector.

[0035] The second analysis module is used to determine the analysis result corresponding to the target object based on the target description vector, the original text vector, the first object sentiment vector, and the preset large language model;

[0036] In one embodiment, the device further includes:

[0037] The second acquisition module is used to acquire the initial description vector to be trained corresponding to multiple candidate information sources, the sample data corresponding to each candidate information source, and the first sample label and the second sample label corresponding to each sample data.

[0038] The first training module is used to train the initial description vector corresponding to each candidate information source based on the preset large language model, the sample data corresponding to each candidate information source, and the first sample label corresponding to each sample data, so as to obtain multiple intermediate candidate description vectors.

[0039] The third analysis module is used to analyze and process the sample data based on the trained sub-model to obtain the second object sentiment vector corresponding to each sample data.

[0040] The second training module is used to retrain the intermediate candidate description vectors corresponding to each candidate information source based on the preset large language model, the sample data, the first sample label corresponding to each sample data and the second object emotion vector, so as to obtain multiple candidate description vectors.

[0041] In one embodiment, the first training module is specifically used to analyze and process the sample data and the initial description vector corresponding to each candidate information source according to a preset large language model to obtain an initial analysis result.

[0042] Based on the initial analysis results and the first sample label, determine the first loss value corresponding to the initial description vector;

[0043] The initial description vector is adjusted based on the first loss value to obtain a new initial description vector, until the first loss value corresponding to the new initial description vector satisfies the first loss condition, thus obtaining multiple intermediate candidate description vectors.

[0044] In one embodiment, the device further includes:

[0045] The first concatenation module is used to concatenate the sample data with the intermediate candidate description vector corresponding to the sample data for each sample data to obtain the first input vector.

[0046] The fourth analysis module is used to compare and analyze each of the first input vectors with the preset historical description text based on the sub-models of the preset large language model to obtain the initial sentiment vector;

[0047] The third training module is used to train the sub-model based on the initial emotion vector and the second sample label to obtain the trained sub-model.

[0048] In one embodiment, the third training module is specifically used to determine a second loss value for the sub-model based on the initial sentiment vector and the second sample label;

[0049] The model parameters of the sub-model are adjusted according to the second loss value until the second loss value meets the second loss condition, thus obtaining the trained sub-model.

[0050] In one embodiment, the sub-model includes historical description text corresponding to the target object; the first analysis module is specifically used to concatenate the target description vector and the original text vector to obtain a second input vector;

[0051] The historical descriptive text and the second input vector are compared and analyzed according to the sub-model. Based on the comparison and analysis results, the influence of the original text vector on the state of the target object is determined, and the first object emotion vector corresponding to the original text vector is obtained.

[0052] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0053] Obtain the original text vector corresponding to the target object; the original text vector includes the target information source corresponding to the original text vector.

[0054] The target description vector corresponding to the target information source is determined according to the correspondence between the candidate information sources and the candidate description vectors; the candidate information sources are multiple information publishing platforms or channels stored in advance, and each candidate information source contains the corresponding candidate description vector, which is a feature supplement for each candidate information source;

[0055] The sub-model based on the preset large language model analyzes and processes the target description vector and the original text vector to obtain the first object sentiment vector of the original text vector;

[0056] The analysis results corresponding to the target object are determined based on the target description vector, the original text vector, the first object sentiment vector, and the preset large language model.

[0057] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0058] Obtain the original text vector corresponding to the target object; the original text vector includes the target information source corresponding to the original text vector.

[0059] The target description vector corresponding to the target information source is determined according to the correspondence between the candidate information sources and the candidate description vectors; the candidate information sources are multiple information publishing platforms or channels stored in advance, and each candidate information source contains the corresponding candidate description vector, which is a feature supplement for each candidate information source;

[0060] The sub-model based on the preset large language model analyzes and processes the target description vector and the original text vector to obtain the first object sentiment vector of the original text vector;

[0061] The analysis results corresponding to the target object are determined based on the target description vector, the original text vector, the first object sentiment vector, and the preset large language model.

[0062] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0063] Obtain the original text vector corresponding to the target object; the original text vector includes the target information source corresponding to the original text vector.

[0064] The target description vector corresponding to the target information source is determined according to the correspondence between the candidate information sources and the candidate description vectors; the candidate information sources are multiple information publishing platforms or channels stored in advance, and each candidate information source contains the corresponding candidate description vector, which is a feature supplement for each candidate information source;

[0065] The sub-model based on the preset large language model analyzes and processes the target description vector and the original text vector to obtain the first object sentiment vector of the original text vector;

[0066] The analysis results corresponding to the target object are determined based on the target description vector, the original text vector, the first object sentiment vector, and the preset large language model.

[0067] The aforementioned object analysis methods, apparatus, computer devices, computer-readable storage media, and computer program products, by determining the target description vector that matches the target information source in the candidate description vectors, can initially determine the characteristics of the target information source to characterize its importance and serve as the basis for analyzing the first object sentiment vector. This allows the sub-model to understand the original text vector at a higher dimension and obtain the first object sentiment vector. Furthermore, by using the target description vector of the target information source and the first object sentiment vector as supplements to the original text vector, the static pre-defined large language model can be assisted in understanding the influence of the original text vector on the current state of the target object. By using the pre-defined large language model as the conditions for analysis and processing the target description vector, the first object sentiment vector, and the original text vector, the accuracy of the analysis results in determining the target object can be improved. Attached Figure Description

[0068] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0069] Figure 1 This is a flowchart illustrating an object analysis method in one embodiment;

[0070] Figure 2 A flowchart illustrating the application of an object analysis method in one embodiment;

[0071] Figure 3 This is a schematic diagram of the process of training candidate description vectors in one embodiment;

[0072] Figure 4This is a schematic diagram of the process of training intermediate candidate description vectors in one embodiment;

[0073] Figure 5 This is a flowchart illustrating an example of training candidate description vectors in one embodiment;

[0074] Figure 6 This is a schematic diagram of the process of training a sub-model in one embodiment;

[0075] Figure 7 This is a schematic diagram of the process of training a sub-model in another embodiment;

[0076] Figure 8 This is a schematic diagram illustrating the process of a sub-model analyzing and processing the original text vector in one embodiment.

[0077] Figure 9 This is a schematic diagram illustrating the training of candidate description vectors using the medical field as an example in one embodiment;

[0078] Figure 10 This is a schematic diagram illustrating the joint training of a sub-model and candidate description vectors, using the medical field as an example in one embodiment.

[0079] Figure 11 This is a schematic diagram illustrating object analysis using the medical field as an example in one embodiment;

[0080] Figure 12 This is a structural block diagram of an object analysis device in one embodiment;

[0081] Figure 13 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0082] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0083] In one embodiment, such as Figure 1 As shown, an object analysis method is provided. This embodiment illustrates the method's application to a terminal. It is understood that this method can also be applied to a server, and further to a system including both a terminal and a server, and is implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0084] Step 102: Obtain the original text vector corresponding to the target object.

[0085] The original text vector includes the target information source corresponding to the original text vector.

[0086] In this embodiment, the target object can be public sentiment regarding the market situation, the market condition, or the effectiveness of a treatment plan. The analysis result corresponding to the target object can be the development trend of public sentiment, the market condition, or the effectiveness of the treatment plan. Furthermore, the original text vector can be news text that can influence public sentiment or the market condition, or a new treatment plan for a specific symptom. In this case, when the original text vector is news text, the target information source can be a newspaper or media outlet; when the original text vector is a treatment plan, the target information source can be a journal or a hospital.

[0087] Taking the target market as an example, when a newspaper or media reports a news article, the user can input text containing the news content, the time of the news report, and the target information source. The terminal then inputs the text into a preset large language model. The encoder or encoding result of the preset large language model encodes the input text, converting it into a vector form to obtain the original text vector. This original text vector represents the quantified news text or treatment plan, etc.

[0088] Step 104: Determine the target description vector corresponding to the target information source based on the correspondence between candidate information sources and candidate description vectors.

[0089] Among them, the candidate information sources are multiple information publishing platforms or channels stored in advance. Each candidate information source contains a corresponding candidate description vector, which is a feature supplement to each candidate information source.

[0090] In this embodiment, candidate description vectors are used to characterize the high-dimensional features of information published by candidate information sources. These high-dimensional features are stored in a pre-defined large language model as vectors representing the high-dimensional features. The candidate description vectors represent the inherent features of the target information source, serving as a feature supplement to the original text vectors. The terminal can pre-store multiple candidate information sources and pre-generated candidate description vectors for each source. There is a correspondence between candidate information sources and candidate description vectors. For example, a candidate information source can be all newspapers or media outlets in a certain region. The candidate description vectors are used to describe the characteristics of a newspaper or media outlet's news publication, such as its immediacy and high authenticity. The characteristics of the newspaper or media outlet's news publication are stored as vector-based candidate description vectors, corresponding to the newspaper or media outlet. This allows for direct acquisition of the target description vector in object analysis applications without repeating the target description vector training process. The original text vectors typically only contain basic semantic information of the original text, such as word frequency and syntactic structure. Candidate descriptive vectors contain richer feature descriptions, such as high-dimensional features like the immediacy and authenticity of the news release. These high-dimensional features help the pre-defined large language model and its sub-models better understand the deeper meaning and contextual information of the text, enabling them to capture subtle differences and complex relationships that might be overlooked in the original text. For example, the authenticity feature of a news story can help the pre-defined large language model and its sub-models distinguish whether a news story is more inclined towards factual reporting or prediction, thereby improving their ability to understand and analyze the development trends of real public sentiment, market conditions, or the corresponding effects of treatment plans.

[0091] After converting the input text into a raw text vector, the terminal determines the target information source in the raw text vector, matches the target information source with the pre-stored candidate information sources, and determines the target description vector corresponding to the target information source from the candidate description vectors according to the correspondence between the candidate information sources and the candidate description vectors.

[0092] Step 106: Based on the sub-model of the preset large language model, analyze and process the target description vector and the original text vector to obtain the first object sentiment vector of the original text vector.

[0093] In this embodiment, the preset large language model includes sub-models. These sub-models analyze the impact of the original text vector on the target object, capturing changes in the current original text vector's influence on the current target object by analyzing historical text and its effects. The terminal inputs the target description vector and the original text vector into the sub-model of the preset large language model. The sub-model analyzes the original text vector and the target description vector to obtain the object's emotional information reflected by the original text vector. Specifically, taking a Transformer large model as an example, the terminal first inputs the original text vector and the target description vector into the Transformer. In the encoder part of the Transformer, a multi-head self-attention mechanism is used to analyze the relationship between the original text vector and the target description vector. Each head focuses on a different subspace, capturing features of different dimensions. Then, a feedforward neural network further processes the vector output by the attention mechanism to extract higher-level features. After each sub-layer, residual connections and layer normalization are used to ensure smooth information transmission and stable gradient flow. The above steps are repeated, stacking to form a multi-layer encoder. Each encoder layer further refines and integrates the features of the original text vector and the target description vector, gaining a deeper understanding of the information layer by layer. Finally, the decoder and output layer generate and output the object's sentiment information. This object sentiment information provides the pre-defined large language model with more granular sentiment information specific to the target object, serving as a further supplement to the analysis basis of the pre-defined large language model.

[0094] In a specific embodiment, taking the market state as an example, the original text vector includes the text vector corresponding to the news content, such as... Figure 2 As shown, this sub-model is a small-scale model (70 billion parameters) with the same architecture as the preset large language model, which has 720 billion parameters. The terminal inputs the text vector corresponding to the news content and the target description vector corresponding to the target information source to which the news content belongs into the sub-model. The sub-model performs preliminary analysis and processing on the text vector and target description vector corresponding to the news content, determining the impact of the current news content on the market state and establishing a preliminary market state result as the first object sentiment vector. This serves as a supplement to the original text vector and target description vector, enabling the preset large language model to understand the current market sentiment and thus better represent the market state.

[0095] Step 108: Determine the analysis results corresponding to the target object based on the target description vector, the original text vector, the first object sentiment vector, and the preset large language model.

[0096] In this embodiment, after obtaining the first object emotion vector output by the sub-model, the terminal concatenates the first object emotion vector with the target description vector and the original text vector. The first object emotion vector further supplements the target description vector and the original text vector. The concatenated target description vector, original text vector, and first object emotion vector are then input into a preset large language model. The preset large language model understands and analyzes the input target description vector, original text vector, and first object emotion vector through a rich multi-dimensional vector space, ultimately obtaining the analysis result corresponding to the target object.

[0097] In a specific embodiment, taking the market state as an example, the original text vector includes the text vector corresponding to the news content, such as... Figure 2 As shown, the input of the pre-defined large language model includes the text vector corresponding to the news content, the target description vector corresponding to the target information source to which the news content belongs, and the preliminary market state output by the sub-model. The pre-defined large language model encodes the text vector, target description vector, and market state corresponding to the input news content into a high-dimensional vector space through its multi-dimensional vector space. The target description vector determines the characteristics of the target information source of the news content, encoding the different credibility, influence, and reporting styles that different target information sources may possess into the pre-defined large language model. Furthermore, the market sentiment reflected in the current market state serves as background information. By supplementing the news content with background information and the characteristics of the target information source, the pre-defined large language model can more accurately predict the impact of the news content on the market state, thereby obtaining the market state analysis results. These analysis results can characterize future market changes and other outcomes.

[0098] In the above object analysis method, by determining the target description vector that matches the target information source in the original text vector from the candidate description vectors, the characteristics of the target information source can be initially determined to characterize the importance of the target information source and serve as the basis for the analysis of the first object sentiment vector. This allows the sub-model to understand the original text vector at a higher dimension and obtain the first object sentiment vector. Furthermore, by using the target description vector of the target information source and the first object sentiment vector as a supplement to the original text vector, the static pre-set large language model can be assisted in understanding the influence of the original text vector on the current state of the target object. By using the pre-set large language model as the conditions for analysis and processing the target description vector, the first object sentiment vector, and the original text vector, the accuracy of the analysis results in determining the target object can be improved.

[0099] In one exemplary embodiment, such as Figure 3 As shown, before step 102, the method further includes steps 302 to 308. Wherein:

[0100] Step 302: Obtain the initial description vector to be trained corresponding to multiple candidate information sources, the sample data corresponding to each candidate information source, and the first sample label and second sample label corresponding to each sample data.

[0101] In this embodiment, the terminal can obtain the initial description vectors to be trained from candidate information sources through random generation. The historical publication information of each candidate information source is used as sample data, and the sample data is pre-labeled to obtain a first sample label, representing the correct analysis result output by the preset large language model after processing each sample data. The first sample label is the correct output of the preset large language model during the training process of the initial description vector, used for training the initial description vector; the second sample label is the object sentiment vector correctly output by the sub-model after processing the sample data, used as the sample label for training the sub-model.

[0102] Step 304: Train the initial description vector corresponding to each candidate information source based on the preset large language model, the sample data corresponding to each candidate information source, and the first sample label corresponding to each sample data to obtain multiple intermediate candidate description vectors.

[0103] In this embodiment, the terminal analyzes and processes the sample data and initial description vectors based on a preset large language model. Utilizing the high-dimensional vector space pre-trained by the preset large language model to understand the sample data and initial description vectors, the inherent parameters of the preset large language model are frozen, ensuring they remain unchanged. Training is then performed only on the initial description vectors input to the preset large language model. The terminal calculates the loss based on the difference between the output of the preset large language model and the first sample label, and corrects and adjusts the initial description vectors using backpropagation to reduce errors, optimize the representational power of the initial description vectors, and completes the initial training of the initial description vectors, obtaining multiple intermediate candidate description vectors.

[0104] Step 306: Analyze and process the sample data based on the trained sub-model to obtain the second object sentiment vector corresponding to each sample data.

[0105] In this embodiment, the terminal inputs the sample data and the intermediate candidate descriptive vectors corresponding to each sample data into the trained sub-model. The sub-model analyzes and processes the sample data and the intermediate candidate descriptive vectors, and analyzes the emotion of the target object that the intermediate candidate descriptive vectors and the sample data can reflect according to the pre-set analysis dimensions to obtain the second object emotion vector corresponding to each sample data.

[0106] Step 308: Based on the preset large language model, sample data, the first sample label and the second object emotion vector corresponding to each sample data, the intermediate candidate description vector corresponding to each candidate information source is retrained to obtain multiple candidate description vectors.

[0107] In this embodiment, the terminal inputs sample data, intermediate candidate descriptive vectors, and a second emotion vector into a preset large language model. The second emotion vector is combined with the sample data and intermediate candidate descriptive vectors. The intermediate candidate descriptive vectors corresponding to each candidate information source are further trained using the output of the preset large language model and the first sample label. The intermediate candidate descriptive vectors are further corrected and adjusted in the manner of backpropagation. Finally, the training of the initial candidate descriptive vectors is completed, resulting in multiple candidate descriptive vectors.

[0108] In an optional embodiment, after the terminal completes the training of the intermediate candidate description vectors and obtains the candidate description vectors, it can also continue to train the sub-model in reverse according to the candidate description vectors. This achieves the joint vector of the initial description vector and the sub-model. The combined loss of the initial description vector and the sub-model is used as the training iteration condition to gradually improve the adaptability of the candidate description vectors to the candidate information sources and the output accuracy of the sub-model.

[0109] In this embodiment, the initial description vector and intermediate candidate description vectors are trained by leveraging the understanding capabilities of the pre-trained pre-defined large language model in high-dimensional vector space. This allows for more accurate adjustment of the initial and intermediate description vectors. When training the intermediate candidate description vectors, a second object's emotion vector is introduced as emotion information during the training process. This enables more comprehensive optimization of the intermediate candidate description vectors, gradually improving the accuracy and effectiveness of the initial description vectors of candidate information sources. Consequently, the adaptability of the candidate description vectors to the candidate information sources is enhanced, which facilitates the subsequent understanding and recording of the importance, expression preferences, and other feature information of different candidate information sources by the pre-defined large language model, thereby improving the accuracy of the pre-defined large language model's analysis and processing.

[0110] In one exemplary embodiment, such as Figure 4 As shown, step 304 includes steps 402 to 406.

[0111] in:

[0112] Step 402: For each candidate information source, analyze and process the sample data and initial description vector corresponding to each candidate information source according to the preset large language model to obtain the initial analysis results.

[0113] In the embodiments of this application, such as Figure 5As shown, the same candidate information source can include multiple sample data. The terminal inputs the sample data of the same information source and the initial description vector of the candidate information source into the preset large language model. The preset large language model can output the corresponding initial analysis result for each sample data. The initial analysis result represents the degree of fit between the initial description vector and the candidate information source. The initial analysis result will be used for subsequent loss calculation and adjustment to optimize the accuracy and effectiveness of the initial description vector.

[0114] Step 404: Determine the first loss value corresponding to the initial description vector based on the initial analysis results and the first sample label.

[0115] In this embodiment, the terminal calculates the loss value between the initial description vector and the first sample label based on the initial analysis result and the corresponding first sample label, using loss value calculation methods such as cross-entropy loss or contrast loss. The first loss value represents the error or deviation between the initial analysis result and the first sample label, measures the accuracy of the initial description vector in representing the features of the information source, and is used to evaluate the fit between the initial description vector and the candidate information source corresponding to the initial description text, so as to determine whether the performance of the initial description vector in describing the candidate information source meets expectations during the iteration process.

[0116] Step 406: Adjust the initial description vector based on the first loss value to obtain a new initial description vector, until the first loss value corresponding to the new initial description vector satisfies the first loss condition, and obtain multiple intermediate candidate description vectors.

[0117] In this embodiment, when the first loss value is lower than a preset loss threshold or the change in the first loss value is lower than a preset loss change threshold, the terminal can use gradient descent optimization to adjust the initial description vector, and continuously calculate the first loss value based on the initial analysis results generated by the adjusted initial description vector and the first sample label. Iteratively calculate the first loss value and optimize the vector for the initial description vector until the first loss value meets the preset loss threshold or the change in the first loss value meets the training condition of the preset loss change threshold. This indicates that the difference between the initial description vector and the first sample label is within the normal range. At this point, the iteration of the initial description vector can be stopped, and multiple intermediate candidate description vectors that have been trained can be obtained.

[0118] In this embodiment, by inputting the sample data of the candidate information source and the initial description vector into the preset large language model, preliminary analysis results are generated. Then, the loss value of the initial description vector is calculated and optimized, thereby gradually adjusting and improving the initial description vector so that it more accurately matches the features of the candidate information source. This ensures that the intermediate candidate description vectors trained in the preliminary training have higher accuracy and effectiveness in representing the features of the candidate information source, thereby improving the prediction performance and analysis capabilities of the sub-model and the preset large language model.

[0119] In one exemplary embodiment, such as Figure 6 As shown, before step 206, it is necessary to train the sub-models of the preset large language model. This method also includes steps 602 to 606. Wherein:

[0120] Step 602: For each sample data, concatenate the sample data with the corresponding intermediate candidate description vector to obtain the first input vector.

[0121] In this embodiment, each target object corresponds to a limited number of candidate information sources. The sample data can be the historical published information of each candidate information source from the existing candidate information sources. Therefore, after training the initial description vector with the sample data to obtain the intermediate candidate description vector, the intermediate candidate description vector has a correlation with the sample data. Then, the terminal concatenates the intermediate candidate description vector with the sample data to obtain the first input vector, which is used as the input of the sub-model.

[0122] Step 604: Based on the sub-models of the preset large language model, compare and analyze each first input vector with the preset historical description text to obtain the initial emotion vector.

[0123] In this embodiment of the application, the sub-model includes pre-stored preset historical description text as the sub-model's understanding of the historical state of the target object. The sub-model compares and analyzes each first output vector with the historical description text to obtain the initial emotion vector corresponding to each first output vector.

[0124] Step 606: Train the sub-model based on the initial emotion vector and the second sample label to obtain the trained sub-model.

[0125] In this embodiment, the label corresponding to each first input vector is the second sample label. The terminal optimizes and adjusts the model parameters of the sub-model based on the difference between the initial emotion vector and the second sample label to obtain the trained sub-model.

[0126] In this embodiment, the first input vector generated by concatenating sample data with intermediate candidate description vectors provides the sub-model with a more informative first input vector. The sub-model generates an initial sentiment vector by comparing and analyzing the first input vector with preset historical description text, and trains it based on the second sample labels, thereby optimizing the sub-model's text sentiment recognition capability, improving the accuracy of the sub-model in performing preliminary state analysis of the target object, and further improving the accuracy of the preset large language model in the state analysis of the target object.

[0127] In one exemplary embodiment, such as Figure 7As shown, step 606 includes steps 702 to 704.

[0128] in:

[0129] Step 702: Determine the second loss value of the sub-model based on the initial sentiment vector and the second sample label.

[0130] In this embodiment, the terminal compares the initial emotion vector with the second sample label and calculates the second loss value of the sub-model to measure the error between the initial emotion vector generated by the sub-model and the actual second sample label, thereby evaluating the output accuracy of the sub-model in the state analysis of the target object and helping the sub-model to guide the adjustment of the sub-model parameters in the state analysis of the target object.

[0131] Step 704: Adjust the model parameters of the sub-model according to the second loss value until the second loss value meets the second loss condition, and obtain the trained sub-model.

[0132] In this embodiment, the terminal adjusts the model parameters of the sub-model based on the calculated second loss value and the gradient descent optimization algorithm, reducing the second loss value between the initial emotion vector output by the sub-model and the second sample label. After completing one iteration of optimization, the terminal continues to iteratively optimize the sub-model until the second loss value between the initial emotion vector output by the optimized sub-model and the second sample label satisfies the second loss condition. The second loss condition can be that the second loss value is lower than a preset loss threshold or that the degree of change in the second loss value is lower than a preset loss change threshold, resulting in a sub-model with higher state prediction accuracy and better generalization ability.

[0133] In this embodiment, by calculating and optimizing the second loss value, the parameters of the sub-model are precisely adjusted, thereby improving the accuracy of the sub-model in generating object emotion vectors, and further improving the accuracy of the preset large language model in target object state analysis.

[0134] In one exemplary embodiment, the sub-model contains historical descriptive text corresponding to the target object; such as Figure 8 As shown, step 106 includes steps 802 to 804. Wherein:

[0135] Step 802: Concatenate the target description vector and the original text vector to obtain the second input vector.

[0136] In this embodiment of the application, during the application process of the preset large language model to perform state analysis on the target object, the terminal concatenates the target description vector corresponding to the target information source to which the original text vector belongs with the original text vector, supplements the original text vector by the feature labeling of the target information source in the high-dimensional space in the target description vector, and inputs the second input vector obtained after concatenation into the sub-model of the preset large language model for preliminary analysis.

[0137] Step 804: Compare and analyze the historical description text and the second input vector according to the sub-model, and based on the comparison and analysis results, determine the influence of the original text vector on the state of the target object, and obtain the first object emotion vector corresponding to the original text vector.

[0138] In this embodiment, the historical description text is the historical record of the sub-model's target object state analysis, including the historical records of the target object's original text vector and target description vector, and the historical object sentiment vector obtained after analyzing the target object's original text vector and target description vector. The terminal compares and analyzes the second input vector with the historical description text in terms of semantics, sentiment, and content, evaluates the semantic and sentiment similarity between the input second input vector and the historical description text, and infers the influence of the current original text vector on the target object's state through comparative analysis, determining the change in the target object's current state. Finally, the sub-model outputs a first object sentiment vector reflecting the target object's current state based on the comparative analysis results, serving as the basis for the subsequent final decision by the preset large language model.

[0139] In this embodiment, the small-scale sub-model focuses on the representation of object emotions and does not need to process a large number of model parameters. Therefore, it can quickly generate the first object emotion vector, reducing the consumption of computing resources. By comparing and analyzing the historical descriptive text and the second input vector, the sub-model can understand the state background of the target object and generate the first object emotion vector. It only analyzes the influence of the original text vector on the state of the target object and also provides key data support for the subsequent state analysis and prediction of the preset large language model.

[0140] In one optional embodiment, an example of an object analysis method is provided, such as Figure 9 As shown, this method can also be applied to the medical field, wherein:

[0141] Relevant indicators in the case (such as Figure 9The transaminase, white blood cell, and creatinine levels shown in the diagram are used as candidate information sources. The reference values ​​corresponding to each relevant indicator are used as sample data. The degree of influence of different indicators and reference values ​​on the diagnostic data is used as the first sample label. The medical diagnostic data of the indicators and reference values ​​in the case are used as the second sample label. The second sample label is the object emotion vector correctly output by the sub-model after processing the sample data. It is used as the sample label for training the sub-model (several transformer decoder models).

[0142] The terminal first uses indicators as candidate information sources, randomly initializes the initial description vectors of each indicator, and processes the initial description vectors through a sub-model to obtain the output result. The inherent parameters of the preset large language model are frozen. Based on the difference between the output result and the first sample label, the loss of the initial description vector of each indicator is calculated. The initial description vectors corresponding to each indicator are trained and adjusted using backpropagation to obtain preliminary training of the initial description vectors, which serve as intermediate candidate description vectors. Finally, the initial description vectors are stored according to the correspondence between indicators and intermediate candidate description vectors, completing the training of the initial description vectors and obtaining candidate description vectors. These candidate description vectors represent the type of influence and interference of different indicators as information sources on the analysis results. Then, as... Figure 10 As shown, the terminal follows the same principle as steps 602 to 606, training the sub-model (the large model to be trained) based on the accuracy of the data processing of the reference values ​​and candidate description vectors corresponding to the relevant indicators. The sub-model is trained by the difference between the user health status vector output by the sub-model and the label of the second sample. At the same time, during the training process of the sub-model, the candidate description vector of the candidate information source can be further optimized based on the output of the preset large language model and the label of the patient's real status.

[0143] This object analysis method is applied in the medical field, such as... Figure 11As shown, the terminal receives the patient's current examination results and past medical examination reports, results, or medical records. First, it analyzes each indicator in the current examination results and its corresponding candidate descriptive vector using a sub-model to obtain the patient's object emotion vector. This object emotion vector represents the initial analysis result obtained by the sub-model after preliminary analysis of the patient's individual information. For example, if the patient's medical diagnosis data includes a history of disease A, and the diseases with a significant impact on the diagnostic data include disease A, disease B, and disease C, then the sub-model output includes an object emotion vector with a higher tendency towards disease A, serving as a supplement to the input vector of the preset large language model. Then, by using the preset language model to input each indicator in the current examination results, its corresponding candidate descriptive vector, and the object emotion vector, the terminal predicts the patient's disease, obtaining the final model prediction result.

[0144] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0145] Based on the same inventive concept, this application also provides an object analysis apparatus for implementing the object analysis method described above. The solution provided by this apparatus is similar to the implementation scheme described in the above method; therefore, the specific limitations in one or more object analysis apparatus embodiments provided below can be found in the limitations of the object analysis method described above, and will not be repeated here.

[0146] In one exemplary embodiment, such as Figure 12 As shown, an object analysis device 1200 is provided, including: a first acquisition module 1201, a determination module 1202, a first analysis module 1203, and a second analysis module 1204, wherein:

[0147] The first acquisition module 1201 is used to acquire the original text vector corresponding to the target object; the original text vector includes the target information source corresponding to the original text vector.

[0148] The determination module 1202 is used to determine the target description vector corresponding to the target information source based on the correspondence between the candidate information sources and the candidate description vectors; the candidate information sources are multiple information publishing platforms or channels stored in advance, each candidate information source contains a corresponding candidate description vector, and the candidate description vector is a feature supplement for each candidate information source;

[0149] The first analysis module 1203 is used to analyze and process the target description vector and the original text vector based on the sub-model of the preset large language model to obtain the first object sentiment vector of the original text vector.

[0150] The second analysis module 1204 is used to determine the analysis results corresponding to the target object based on the target description vector, the original text vector, the first object sentiment vector and the preset large language model.

[0151] In one embodiment, the device 1200 further includes:

[0152] The second acquisition module is used to acquire the initial description vector to be trained corresponding to multiple candidate information sources, the sample data corresponding to each candidate information source, and the first sample label and the second sample label corresponding to each sample data.

[0153] The first training module is used to train the initial description vector corresponding to each candidate information source based on the preset large language model, the sample data corresponding to each candidate information source, and the first sample label corresponding to each sample data, so as to obtain multiple intermediate candidate description vectors.

[0154] The third analysis module is used to analyze and process the sample data based on the trained sub-model to obtain the second object sentiment vector corresponding to each sample data.

[0155] The second training module is used to retrain the intermediate candidate description vectors corresponding to each candidate information source based on the preset large language model, sample data, the first sample label and the second object emotion vector corresponding to each sample data, so as to obtain multiple candidate description vectors.

[0156] In one embodiment, the first training module is specifically used to analyze and process the sample data and initial description vector corresponding to each candidate information source according to a preset large language model to obtain the initial analysis result.

[0157] Determine the first loss value corresponding to the initial description vector based on the initial analysis results and the first sample label;

[0158] The initial description vector is adjusted based on the first loss value to obtain a new initial description vector. This process continues until the first loss value corresponding to the new initial description vector satisfies the first loss condition, resulting in multiple intermediate candidate description vectors.

[0159] In one embodiment, the device 1200 further includes:

[0160] The first concatenation module is used to concatenate the sample data with the corresponding intermediate candidate description vector for each sample data to obtain the first input vector.

[0161] The fourth analysis module is used to compare and analyze each first input vector with the preset historical description text based on the sub-models of the preset large language model to obtain the initial sentiment vector;

[0162] The third training module is used to train the sub-model based on the initial sentiment vector and the second sample labels to obtain the trained sub-model.

[0163] In one embodiment, the third training module is specifically used to determine a second loss value for the sub-model based on the initial sentiment vector and the second sample label;

[0164] The model parameters of the sub-model are adjusted based on the second loss value until the second loss value meets the second loss condition, thus obtaining the trained sub-model.

[0165] In one embodiment, the sub-model contains historical description text corresponding to the target object; the first analysis module is specifically used to concatenate the target description vector and the original text vector to obtain the second input vector;

[0166] The sub-model compares and analyzes the historical descriptive text and the second input vector, and based on the comparison analysis results, determines the influence of the original text vector on the state of the target object, thus obtaining the first object emotion vector corresponding to the original text vector.

[0167] Each module in the aforementioned object analysis device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0168] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 13As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores historical descriptive text. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When executed by the processor, the computer program implements an object analysis method.

[0169] Those skilled in the art will understand that Figure 13 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0170] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0171] Obtain the original text vector corresponding to the target object; the original text vector includes the target information source corresponding to the original text vector.

[0172] The target description vector corresponding to the target information source is determined based on the correspondence between candidate information sources and candidate description vectors;

[0173] The sub-model based on the pre-defined large language model analyzes and processes the target description vector and the original text vector to obtain the first object sentiment vector of the original text vector.

[0174] The analysis results corresponding to the target object are determined based on the target description vector, the original text vector, the first object sentiment vector, and the pre-set large language model.

[0175] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0176] Obtain the initial description vector to be trained corresponding to multiple candidate information sources, the sample data corresponding to each candidate information source, and the first sample label and second sample label corresponding to each sample data;

[0177] Based on the pre-set large language model, the sample data corresponding to each candidate information source, and the first sample label corresponding to each sample data, the initial description vector corresponding to each candidate information source is trained to obtain multiple intermediate candidate description vectors.

[0178] Based on the trained sub-model, the sample data is analyzed and processed to obtain the second object sentiment vector corresponding to each sample data.

[0179] Based on the pre-set large language model, sample data, the first sample label and the second object emotion vector corresponding to each sample data, the intermediate candidate description vectors corresponding to each candidate information source are retrained to obtain multiple candidate description vectors.

[0180] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0181] For each candidate information source, the sample data and initial description vector corresponding to each candidate information source are analyzed and processed according to the preset large language model to obtain the initial analysis results.

[0182] Determine the first loss value corresponding to the initial description vector based on the initial analysis results and the first sample label;

[0183] The initial description vector is adjusted based on the first loss value to obtain a new initial description vector. This process continues until the first loss value corresponding to the new initial description vector satisfies the first loss condition, resulting in multiple intermediate candidate description vectors.

[0184] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0185] For each sample data, the sample data and the corresponding intermediate candidate description vector are concatenated to obtain the first input vector;

[0186] Based on the sub-models of the preset large language model, each first input vector is compared and analyzed with the preset historical description text to obtain the initial emotion vector;

[0187] The sub-model is trained based on the initial sentiment vector and the second sample labels to obtain the trained sub-model.

[0188] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0189] The second loss value of the sub-model is determined based on the initial sentiment vector and the second sample labels;

[0190] The model parameters of the sub-model are adjusted based on the second loss value until the second loss value meets the second loss condition, thus obtaining the trained sub-model.

[0191] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0192] The target description vector and the original text vector are concatenated to obtain the second input vector;

[0193] The sub-model compares and analyzes the historical descriptive text and the second input vector, and based on the comparison analysis results, determines the influence of the original text vector on the state of the target object, thus obtaining the first object emotion vector corresponding to the original text vector.

[0194] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0195] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0196] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0197] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0198] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0199] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. An object analysis method, characterized in that, The method includes: Obtain the original text vector corresponding to the target object; the original text vector corresponds to the target information source. The target description vector corresponding to the target information source is determined according to the correspondence between the candidate information sources and the candidate description vectors; the candidate information sources are multiple information publishing platforms or channels stored in advance, and each candidate information source contains the corresponding candidate description vector, which is a feature supplement for each candidate information source; The sub-model based on the preset large language model analyzes and processes the target description vector and the original text vector to obtain the first object sentiment vector of the original text vector; The analysis results corresponding to the target object are determined based on the target description vector, the original text vector, the first object sentiment vector, and the preset large language model. Before obtaining the original text vector corresponding to the target object, the method further includes: Obtain the initial description vector to be trained corresponding to multiple candidate information sources, the sample data corresponding to each candidate information source, and the first sample label and the second sample label corresponding to each sample data; Based on the preset large language model, the sample data corresponding to each candidate information source and the first sample label corresponding to each sample data, the initial description vector corresponding to each candidate information source is trained to obtain multiple intermediate candidate description vectors. The sample data is analyzed and processed based on the trained sub-model to obtain the second object sentiment vector corresponding to each sample data. Based on the preset large language model, the sample data, the first sample label corresponding to each sample data and the second object emotion vector, the intermediate candidate description vector corresponding to each candidate information source is retrained to obtain multiple candidate description vectors.

2. The method according to claim 1, characterized in that, The method involves training the initial description vectors corresponding to each candidate information source based on a preset large language model, sample data corresponding to each candidate information source, and first sample labels corresponding to each sample data, to obtain multiple intermediate candidate description vectors, including: For each candidate information source, the sample data corresponding to each candidate information source and the initial description vector are analyzed and processed according to the preset large language model to obtain the initial analysis results; Based on the initial analysis results and the first sample label, determine the first loss value corresponding to the initial description vector; The initial description vector is adjusted based on the first loss value to obtain a new initial description vector, until the first loss value corresponding to the new initial description vector satisfies the first loss condition, thus obtaining multiple intermediate candidate description vectors.

3. The method according to claim 1, characterized in that, Before analyzing and processing the sample data based on the trained sub-model to obtain the second object sentiment vector corresponding to each sample data, the method further includes: For each sample data, the sample data is concatenated with the intermediate candidate description vector corresponding to the sample data to obtain the first input vector; Based on the sub-models of the preset large language model, each of the first input vectors is compared and analyzed with the preset historical description text to obtain the initial emotion vector; The sub-model is trained based on the initial emotion vector and the second sample label to obtain the trained sub-model.

4. The method according to claim 3, characterized in that, The process of training the sub-model based on the initial emotion vector and the second sample label to obtain the trained sub-model includes: The second loss value of the sub-model is determined based on the initial sentiment vector and the second sample label; The model parameters of the sub-model are adjusted according to the second loss value until the second loss value meets the second loss condition, thus obtaining the trained sub-model.

5. The method according to claim 1, characterized in that, The sub-model contains historical descriptive text corresponding to the target object; the sub-model based on the preset large language model analyzes and processes the target description vector and the original text vector to obtain a first object sentiment vector of the original text vector, including: The target description vector and the original text vector are concatenated to obtain the second input vector; The historical descriptive text and the second input vector are compared and analyzed according to the sub-model. Based on the comparison and analysis results, the influence of the original text vector on the state of the target object is determined, and the first object emotion vector corresponding to the original text vector is obtained.

6. An object analysis device, characterized in that, The device includes: The first acquisition module is used to acquire the original text vector corresponding to the target object; the original text vector corresponds to the target information source. The determining module is used to determine the target description vector corresponding to the target information source based on the correspondence between candidate information sources and candidate description vectors; the candidate information sources are multiple information publishing platforms or channels stored in advance, each candidate information source contains the corresponding candidate description vector, and the candidate description vector is a feature supplement for each candidate information source; The first analysis module is used to analyze and process the target description vector and the original text vector based on the sub-model of the preset large language model to obtain the first object sentiment vector of the original text vector. The second analysis module is used to determine the analysis result corresponding to the target object based on the target description vector, the original text vector, the first object sentiment vector, and the preset large language model; The device further includes: The second acquisition module is used to acquire the initial description vector to be trained corresponding to multiple candidate information sources, the sample data corresponding to each candidate information source, and the first sample label and the second sample label corresponding to each sample data. The first training module is used to train the initial description vector corresponding to each candidate information source based on the preset large language model, the sample data corresponding to each candidate information source, and the first sample label corresponding to each sample data, so as to obtain multiple intermediate candidate description vectors. The third analysis module is used to analyze and process the sample data based on the trained sub-model to obtain the second object sentiment vector corresponding to each sample data. The second training module is used to retrain the intermediate candidate description vectors corresponding to each candidate information source based on the preset large language model, the sample data, the first sample label corresponding to each sample data and the second object emotion vector, so as to obtain multiple candidate description vectors.

7. The apparatus according to claim 6, characterized in that, The first training module is specifically used to analyze and process the sample data and the initial description vector corresponding to each candidate information source according to a preset large language model to obtain the initial analysis result; Based on the initial analysis results and the first sample label, determine the first loss value corresponding to the initial description vector; The initial description vector is adjusted based on the first loss value to obtain a new initial description vector, until the first loss value corresponding to the new initial description vector satisfies the first loss condition, thus obtaining multiple intermediate candidate description vectors.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 5.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.

Images