Large language model enhanced text mapping processing method and system for personality assessment

Abstract

The present invention discloses a large language model enhanced text mapping processing method and system for personality assessment, and the method includes: S1, constructing an association model sample database; constructing a BERT model to obtain the text features of user social text data and then average pooling to obtain an initial vector; S2, first performing social text data splicing processing, extracting and embedding processing through the Llama3 model, and average pooling processing to obtain a user integration vector; S3, constructing a fusion model including three fully connected layers and an attention mechanism to perform addition fusion to obtain the final user vector; S4, using the GPT‑4 model and the BERT model to interpret and encode personality labels, and the target label comparison learning model to perform comparison learning model training. The present invention uses a model sample database to train model processing and relationships such as the BERT model, the Llama3 model, the fusion model, and the target label comparison learning model, which significantly improves the learning accuracy and efficiency of the user vector and personality label relationship model.

Description

Technical Field

[0001] The present invention relates to the field of deep learning models, and in particular to a large language model enhanced text mapping processing method and system for personality assessment. Background Art

[0002] In the field of civil aviation, people, machines, and the environment are the key factors in the risks of civil aviation operations. At present, there are many risk assessments for machines and the environment. There are also many academic studies on human factors, such as "Construction and Verification of Psychological Competency Characteristics Model of Civil Aviation Pilots in Crisis Situations" (Author: Liu Chuanjian), which points out that the main components of the psychological competence of civil aviation pilots in crisis situations include psychological resilience, sharpness and decisiveness, calmness, self-efficacy, strong will and courage to take responsibility, and also constructs the psychological competence assessment dimension and system. For example, the "Human Factors in Flight" of Nanjing University of Aeronautics and Astronautics statistics that aircraft accidents involving human factors can be as high as 90%. In civil aviation flights, there are crew members, ground service personnel, air traffic control personnel, etc. In order to better ensure flight safety, personnel entering the civil aviation field need to be evaluated in terms of personality and other aspects. MBTI (Myers-Briggs Type Indicator, abbreviated as MBTI) is a widely used personality classification system. It is a personality type theory model jointly developed by American writer Isabel Briggs Myers and her mother Katherine Cook Briggs. It divides individuals into sixteen categories based on four dimensions. More importantly, MBTI has rich sample data, which is extremely important for deep learning. The MBTI test can help us better understand people's personality qualities.

[0003] At present, the evaluation of civil aviation staff is mainly based on scaled questions and answers. This method can only reflect the psychological state of the staff when answering questions. There are also large deviations in the scoring method, which cannot better reflect the quality of the staff. "Construction and Verification of Psychological Competency Characteristics Model of Civil Aviation Pilots in Crisis Situations" adopts text encoding under big data. Existing implementation methods (such as: Keh, Sedrick Scott & Cheng, I-Tsun. (2019). Myers-Briggs Personality Classification and Personality-Specific Language Generation Using Pre-trained Language Models. 10.48550 / arXiv.1907.06333) use the Bert model for text encoding (such as: Lee J, Toutanova K. Pre-training of deep bidirectional transformers for language understanding [J]. arxiv preprint arxiv: 1810.04805, 2018, 3 (8)), but the Bert model has limited representation capabilities for multiple long texts when encoding text, and its performance in capturing the relationship between the user representation vector and the personality label vector is poor. These methods have some shortcomings and problems. First, these methods use a small model such as Bert for text encoding, which has limited representation capabilities for multiple long texts. Second, these methods do not effectively capture the relationship between user representation vectors and personality label vectors (the traditional method is to extract static features from the text, encode the text information to obtain user vectors, and then map them to personality labels), resulting in poor performance. Some research schemes use psychological statistical feature similarity or post embedding similarity to measure whether there is a correlation between posts, and then use graph neural networks to construct the post text set topology and perform feature fusion between post questions. Traditional methods rely solely on small model semantic encoders to extract semantic features from multiple long texts. Deep learning methods often produce low-quality user vectors, and the relationship between these vectors and MBTI labels has not been fully determined. Based on this, it is urgent to study a model method suitable for personnel to conduct personality quality assessment based on big data, to achieve effective guidance on the adaptation and selection of human quality and ability, which is also an important direction for civil aviation safety operation research. Summary of the invention

[0004] The purpose of the present invention is to solve the technical problems pointed out by the background technology, and to provide a large language model enhanced text mapping processing method and system for personality assessment, which uses a model sample database to train the model method, obtains text features from a user social text dataset through the BERT model and performs average pooling to obtain the user's initial vector, performs splicing processing on the same user, and then uses the Llama3 model to extract and embed to obtain a synthesized user integrated vector, thereby enhancing the ability to expand text and providing unique semantic insights; then the fusion model uses three fully connected layers and an attention mechanism to fuse to obtain the final user vector; the personality label is interpreted from K dimensions through a GPT-4 model and encoded using a BERT model, and a comparative learning model is trained using a target label comparative learning model, thereby realizing the integration of user vector and personality label encoding into comparative learning, and significantly improving the learning accuracy and efficiency of the user vector and personality label relationship model.

[0005] The purpose of the present invention is achieved through the following technical solutions:

[0006] A large language model enhanced text mapping processing method for personality assessment, the method comprising:

[0007] S1. Build a model sample database that associates user social text datasets and personality labels. The user social text dataset contains M users, where user m contains N social text data p i , p i Represents the i-th social text data of the user; build the BERT model and input the user social text dataset, and then obtain the text feature h of the user social text data i , all the user's text features h i Perform average pooling to obtain the initial vector u corresponding to the user;

[0008] S2: The user's N social text data p i The concatenation and merging process is performed, and then the Llama3 model is used for extraction and embedding, followed by average pooling to obtain a synthetic user integration vector;

[0009] S3. Construct a fusion model including three fully connected layers and an attention mechanism. The fusion model first transforms the initial vector u and the user integration vector U llama After the conversion process, three fully connected layers are used to map the vector to the attention mechanism and generate the vector H llama ; Then vector H llama Add and fuse with the initial vector u to get the final user vector H of user m p ;

[0010] S4. Use the GPT-4 model to interpret the personality labels of the model sample database from K dimensions and then use the BERT model to encode the personality label encoding, and then build the target label comparison learning model to input the user vector H p , personality label encoding for comparative learning model training.

[0011] Preferably, the present invention also includes the following method:

[0012] S5: Collect the user's corresponding social text data, and process it in sequence according to steps S2 to S3 to obtain the user vector H p , and then input the target label to compare the learning model to obtain the final predicted personality label.

[0013] Preferably, in step S4, each label code contains L+1 feature values, and the jth label code feature set is Indicates that the jth label corresponds to the Lth eigenvalue, and the target label comparison learning model classifies the eigenvalue corresponding to the target user m as l j+ , classify the feature values ​​that do not correspond to the target user m as l j- , thereby constructing the contrast loss function L of the model cl as follows:

[0014] Where sim(u,l j+ ) represents the embedding vectors u and l j+ The cosine similarity between sim(u,l j- ) represents the embedding vectors u and l j- The cosine similarity between them, τ is the parameter for adjusting the similarity sensitivity, and T represents the total number of personality labels.

[0015] Preferably, the target label contrast learning model also adopts a focal loss function L fl ,

[0016]

[0017] Where α represents the weight factor, γ represents the focusing parameter of the sample weight, represents the predicted classification probability of the model parameters θ, social text data i, and label j, y ji represents the true classification probability of social text data i and label j, and V represents the total number of samples in the model sample database;

[0018] The target label contrastive learning model uses the following loss constraint that combines the contrastive loss function with the focal loss function: L = L fl +λL cl , where λ represents a hyperparameter.

[0019] Preferably, in step S1, the token expression of the CLS position encoded by the BERT model is used as the social text data p i The text feature h i ; The set of all user text features is [h1, h2, …h i …,h N ], the initial vector u is the set of all user text features [h1, h2, …h i …,h N ] is obtained by average pooling processing; the BERT model takes the embedding of the CLS position as the label embedding when obtaining the personality label encoding.

[0020] Preferably, in step S2, N social text data p i The Llama3 model includes 32 layers, and the embedding hyperparameter dm is set. The embedding processing from the 33rd-dm layer to the 32nd layer is extracted and average pooled to obtain the user integration vector U llama .

[0021] Preferably, in step S3, the fusion model integrates the user vector U llama Using the transformation matrix W o1 , W o2 The user integration vector U llama After conversion, the attention mechanism of the fusion model has Q query vector space, K key vector space and V value vector space. The mapping relationship expression is as follows:

[0022] Q=uW Q

[0023] K=(U llama W o1 )W K

[0024] V=(U llama W o2 )W V

[0025] W Q , W K and W V are the projection matrices respectively.

[0026] As a preference, the dimensions of the Q query vector space and the K key vector space of the attention mechanism in the fusion model are both d k , the attention mechanism of the fusion model generates the vector H llama The expression is as follows:

[0027] Softmax represents a normalization function.

[0028] A large language model enhanced text mapping processing system for personality assessment includes a model sample database, a BERT model, a Llama3 model, a fusion model, a GPT-4 model and a target label comparison learning model. The model sample database stores user social text data sets and personality labels in sequence according to users. The model sample database contains M users, where user m contains N social text data sets p. i , p i represents the i-th social text data of the user; the BERT model is used to input the user social text data set and extract features from the user social text data to obtain the text feature H corresponding to the user m i , all text features h of user m i The average pooling process is performed to obtain the initial vector u corresponding to the user; the Llama3 model is used to process the N social text data p of user m. i After the splicing process, the data is extracted and embedded, and then average pooled to obtain the synthesized user integration vector U llama The fusion model includes three fully connected layers and an attention mechanism. The fusion model first transforms the initial vector u and the user integration vector U llama After the conversion, three fully connected layers are used to map the vector to the attention mechanism and generate the vector H llama , then vector H llama Add and fuse with the initial vector u to get the final user vector H of user m p ; The GPT-4 model encodes the personality labels of the model sample database from K dimensions; the target label comparison learning model inputs the user vector H p , personality label encoding for comparative learning model training.

[0029] Preferably, the large language model enhanced text mapping processing system for personality assessment of the present invention further includes a new user data collection module, which is used to collect user social text data corresponding to the user and process it through the BERT model, Llama3 model, and fusion model to obtain a user vector H p The target label contrast learning model is based on the input user vector H p Get the final predicted personality label.

[0030] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0031] (1) The present invention uses a model sample database to train the model method, obtains text features from a user social text dataset through a BERT model, and performs average pooling to obtain the user's initial vector, performs splicing on the same user, and then uses a Llama3 model to extract and embed to obtain a synthetic user integration vector, thereby enhancing the ability to expand text and providing unique semantic insights; then the fusion model uses three fully connected layers and an attention mechanism to fuse to obtain the final user vector; the personality label is interpreted from K dimensions through a GPT-4 model and encoded using a BERT model, and a target label contrast learning model is used to train a contrast learning model, thereby realizing the integration of user vector and personality label encoding into contrast learning, and significantly improving the learning accuracy and efficiency of the user vector and personality label relationship model.

[0032] (2) The large language Llama3 model of the present invention enhances the user vector representation and provides rich semantic embedding as a supplement to text features; the fusion model includes three fully connected layers and an attention mechanism, which fully utilizes the initial vector of the BERT model and the user integrated vector output by the Llama3 model for conversion, fully connected layer processing, attention mechanism processing, and addition and fusion with the initial vector to obtain the final user vector, which richly and deeply obtains various feature data of the user social text dataset, providing key data support for subsequent evaluation.

[0033] (3) The present invention uses the GPT-4 model to interpret personality labels from K dimensions to obtain rich data content, which facilitates the BERT model to perform comprehensive and accurate label encoding, provides data support for the comprehensive comparison of subsequent target label comparison learning models, and significantly improves the comparison accuracy and quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 A method flow chart of a large language model enhanced text mapping processing method for personality assessment according to the present invention;

[0035] Figure 2 Schematic diagram of the principle of the large language model enhanced text mapping processing method for personality assessment in the embodiment;

[0036] Figure 3 This is a principle structural block diagram of the large language model enhanced text mapping processing system for personality assessment of the present invention. DETAILED DESCRIPTION

[0037] The present invention is further described in detail below in conjunction with embodiments:

[0038] Example

[0039] like Figure 1 , Figure 2As shown, a large language model enhanced text mapping processing method for personality assessment, the method includes:

[0040] S1. Construct a model sample database that associates user social text datasets with personality labels. The data of the model sample database in this embodiment comes from the MBTI personality datasets of Kaggle and Pandora or from other sample data. The Kaggle dataset comes from PersonalityCafe, and each entry lists a person's four-letter MBTI type and an excerpt of their most recent 50 post texts. The Pandora dataset comes from Reddit and contains 9067 users (the number of post texts for each user ranges from L tens to hundreds). Both the Kaggle dataset and the Pandora dataset contain label data. The user social text dataset contains M users, where user m contains N social text data p i , p i Represents the user’s i-th social text data. Build a BERT model and input the user’s social text dataset, then obtain the text feature h of the user’s social text data i , all the user's text features h i The average pooling process is performed to obtain the initial vector u corresponding to the user. The BERT model encodes all social text data of user m (including N social text data) respectively, and then obtains the initial vector u through the average pooling method. In some embodiments, the token expression of the CLS position encoded by the BERT model is used as the social text data p i The text feature h i The set of all user text features is [h1, h2, …h i …h N ], the initial vector u is the set of all user text features [h1, h2, …h i …,h N ] is obtained by average pooling. The BERT model uses the embedding of the CLS position as the label embedding when obtaining the personality label encoding; the BERT model is used as a text encoder for social text data p i Encode, then extract the token at the [CLS] position, and then use the token at the [CLS] position as the social text data p i The features are used to represent the text features h i .

[0041] S2: The user's N social text data p i Perform connection merging processing.

[0042] In some embodiments, N social text data p iThe Llama3 model includes 32 layers, and the embedding hyperparameter dm is set (the embedding hyperparameter dm is the last number of layers selected by the Llama3 model). The embedding processing from the 33-dm layer to the 32nd layer is extracted and average pooled to obtain the user integration vector U llama If the embedding hyperparameter dm is set to 5, the Llama3 model extracts layers 28 to 32 and performs average pooling within and between layers to capture richer semantic details. llama =mean([U 32-dm+1 , U 32-dm+2 , …, U 32 ]).

[0043] S3. Construct a fusion model including three fully connected layers and an attention mechanism. The fusion model first transforms the initial vector u and the user integration vector U llama The conversion process is then performed, and then three fully connected layers are used to map the vector to the attention mechanism (preferably the cross attention mechanism) and generate the vector H llama Then the vector H llama Add and fuse with the initial vector u to get the final user vector H of user m p .

[0044] In some embodiments, the fusion model integrates the user vector U llama Using the transformation matrix W o1 , W o2 The user integration vector U llama After conversion, the attention mechanism of the fusion model has Q query vector space, K key vector space and V value vector space. The mapping relationship expression is as follows:

[0045] Q=uW Q

[0046] K=(U llama W o1 )W K

[0047] V=(U llama W o2 )W V

[0048] W Q , W K and W V are the projection matrices respectively.

[0049] The dimensions of the Q query vector space and K key vector space of the attention mechanism in the fusion model are both d k (d kis the set scaled attention size), the attention mechanism of the fusion model generates the vector H llama The expression is as follows:

[0050] Softmax represents a normalization function.

[0051] S4. Use the GPT-4 model to interpret the personality labels of the model sample database from K dimensions (K can be selected from the same four dimensions as the Myers-Briggs Type Indicator MBTI, or from the three dimensions of type definition, theme tendency and expression) (preferably, the maximum pooling method is used to maximize the retention of MBTI label interpretation after interpretation) and then use the BERT model to encode the personality label encoding, and then build the target label comparison learning model to input the user vector H p , personality label encoding for comparative learning model training.

[0052] In some embodiments, each label encoding contains L+1 feature values, and the jth label encoding feature set is Indicates that the jth label corresponds to the Lth eigenvalue, and the target label comparison learning model classifies the eigenvalue corresponding to the target user m as l j+ , classify the feature values ​​that do not correspond to the target user m as l j- , thereby constructing the contrast loss function L of the model cl as follows:

[0053] Where sim(u,l j+ ) represents the embedding vectors u and l j+ The cosine similarity between sim(u,l j- ) represents the embedding vectors u and l j- The cosine similarity between them, τ is the parameter for adjusting the similarity sensitivity, T represents the total number of personality labels, T=16.

[0054] Preferably, the target label contrast learning model also uses a focal loss function L fl ,

[0055]

[0056] Where α represents the weight factor, γ represents the focusing parameter of the sample weight, represents the predicted classification probability of the model parameters θ, social text data i, and label j, y ji represents the true classification probability of social text data i and label j, and V represents the total number of samples in the model sample database.

[0057] The target label contrastive learning model uses the following loss constraint that combines the contrastive loss function with the focal loss function: L = L fl +λL cl , where λ represents a hyperparameter (adjusting the relative importance of each loss in the objective function).

[0058] S5. Collect user social text data corresponding to the user. In the field of civil aviation, the source of user social text data includes social text data of civil aviation staff (such as extracted network posts or blog articles or social dialogue data in civil aviation work) or social text data extracted through audio or video conversion. According to steps S2 to S3 of the present invention, the user vector H is obtained by sequentially processing. p , and then input the target label to compare the learning model to obtain the final predicted personality label.

[0059] The present invention achieved 10.09% and 10.19% performance improvement on the benchmark dataset compared with the traditional TrigNet method (TrigNet is an important part of the Path-based DeepNetwork (PDN) model, which is used to calculate the user's interest in each item that has been historically interacted with. It uses the user's characteristics, the characteristics of the historically interacted items, and the interaction behavior as input, and uses a multi-layer perceptron (MLP) to calculate the user's preference for each interactive item, thereby obtaining the user's multiple interest scores). By effectively utilizing the post-concatenation representation, it achieved 9.03% and 7.12% performance improvement over D-DGCN (DoubleDynamic Graph Convolutional Network, which simultaneously processes the original graph and the dual graph to fuse more relationship information, thereby learning better entity representations. D-DGCN may focus more on the processing and analysis of dynamic graph data). The large language model of the present invention performs splicing and fusion processing of all social text data of users, and also achieves significant improvements of 28.35% and 22%.

[0060] The present invention also uses the Kaggle dataset for ablation research. When the Llama3 model is removed, the overall performance drops by 16.8%; when the GPT-4 model is removed to explain the personality labels of the model sample database from K dimensions, the performance drops by 6.97%. Figure 2 The technical principle significantly enhances the mapping process from user text data to personality labels.

[0061] like Figure 3As shown, a large language model enhanced text mapping processing system for personality assessment includes a model sample database, a BERT model, a Llama3 model, a fusion model, a GPT-4 model and a target label comparison learning model. The model sample database stores user social text data sets and personality labels in sequence according to users. The model sample database contains M users, where user m contains N social text data sets p i , p i Represents the i-th social text data of the user. The BERT model is used to input the user social text data set and extract features from the user social text data to obtain the text feature h corresponding to user m. i , all text features h of user m i The Llama3 model is used to average pool the N social text data p of user m. i After the splicing process, the data is extracted and embedded, and then average pooled to obtain the synthesized user integration vector U llama The fusion model includes three fully connected layers and an attention mechanism. The fusion model first transforms the initial vector u and the user integration vector U llama After the conversion, three fully connected layers are used to map the vector to the attention mechanism and generate the vector H llama , then vector H llama Add and fuse with the initial vector u to get the final user vector H of user m p The GPT-4 model encodes the personality labels of the model sample database from K dimensions. The target label contrast learning model inputs the user vector H p , personality label encoding for comparative learning model training.

[0062] The large language model enhanced text mapping processing system for personality assessment of the present invention also includes a new user data collection module, which is used to collect user social text data corresponding to the user and process it through the BERT model, Llama3 model, and fusion model to obtain a user vector H p The target label contrast learning model is based on the input user vector H p Get the final predicted personality label.

[0063] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A large language model enhanced text mapping processing method for personality assessment, characterized by: The methods include: S1. Build a model sample database that associates user social text datasets and personality labels. The user social text dataset contains M users, where user m contains N social text data p i , p i Represents the i-th social text data of the user; build the BERT model and input the user social text dataset, and then obtain the text feature h of the user social text data i , all the user's text features h i Perform average pooling to obtain the initial vector u corresponding to the user; S2: The user's N social text data p i The concatenation and merging process is performed, and then the Llama3 model is used for extraction and embedding, followed by average pooling to obtain a synthetic user integration vector; S3. Construct a fusion model including three fully connected layers and an attention mechanism. The fusion model first transforms the initial vector u and the user integration vector U llama After the conversion, three fully connected layers are used to map the vector to the attention mechanism and generate the vector H llama ; Then vector H llama Add and fuse with the initial vector u to get the final user vector H of user m p ; S4. Use the GPT-4 model to interpret the personality labels of the model sample database from K dimensions and then use the BERT model to encode the personality label encoding, and then build the target label comparison learning model to input the user vector H p , personality label encoding for comparative learning model training.

2. The large language model enhanced text mapping processing method for personality assessment according to claim 1, characterized in that: Also includes the following methods: S5: Collect the user's corresponding social text data, and process it in sequence according to steps S2 to S3 to obtain the user vector H p , and then input the target label to compare the learning model to obtain the final predicted personality label.

3. The large language model enhanced text mapping processing method for personality assessment according to claim 1 or 2, characterized in that: In step S4, each label code contains L+1 feature values, and the jth label code feature set is Indicates that the jth label corresponds to the Lth eigenvalue, and the target label comparison learning model classifies the eigenvalue corresponding to the target user m as l j+ , classify the feature values ​​that do not correspond to the target user m as l j- , thereby constructing the contrast loss function L of the model cl as follows: Where sim(u,l j+ ) represents the embedding vectors u and l j+ The cosine similarity between sim(u,l j- ) represents the embedding vectors u and l j- The cosine similarity between them, τ is the parameter for adjusting the similarity sensitivity, and T represents the total number of personality labels.

4. The large language model enhanced text mapping processing method for personality assessment according to claim 3 is characterized in that: The target label contrast learning model also uses a focal loss function L fl , Where α represents the weight factor, γ represents the focusing parameter of the sample weight, represents the predicted classification probability of the model parameters θ, social text data i, and label j, y ji represents the true classification probability of social text data i and label j, and V represents the total number of samples in the model sample database; The target label contrastive learning model uses the following loss constraint that combines the contrastive loss function with the focal loss function: L = L f x+λL cl , where λ represents a hyperparameter.

5. The large language model enhanced text mapping processing method for personality assessment according to claim 1 is characterized in that: in step S1, the token expression of the CLS position encoded by the BERT model is used as the social text data p i The text feature h i ; The set of all user text features is [h1, h2, …h l …,h N ], the initial vector u is the set of all user text features [h1, h2, …h l …,h N ] is obtained by average pooling processing; the BERT model takes the embedding of the CLS position as the label embedding when obtaining the personality label encoding.

6. The large language model enhanced text mapping processing method for personality assessment according to claim 1, characterized in that: In step S2, N social text data p i The Llama3 model includes 32 layers, and the embedding hyperparameter dm is set. The embedding processing from the 33rd-dm layer to the 32nd layer is extracted and average pooled to obtain the user integration vector U llama .

7. The large language model enhanced text mapping processing method for personality assessment according to claim 1, characterized in that: In step S3, the fusion model integrates the user vector U llama Using the transformation matrix W o1 , W o2 The user integration vector U llama After conversion, the attention mechanism of the fusion model has Q query vector space, K key vector space and V value vector space. The mapping relationship expression is as follows: Q=uW Q K=(U llama W o1 )W K V=(U llama W o2 )W V W Q , W K and W V are the projection matrices respectively.

8. The large language model enhanced text mapping processing method for personality assessment according to claim 7, characterized in that: The dimensions of the Q query vector space and K key vector space of the attention mechanism in the fusion model are both d k , the attention mechanism of the fusion model generates the vector H llama The expression is as follows: Softmax represents a normalization function.

9. A large language model enhanced text mapping processing system for personality assessment, characterized by: It includes a model sample database, a BERT model, a Llama3 model, a fusion model, a GPT-4 model and a target label comparison learning model. The model sample database stores user social text data sets and personality labels in sequence according to users. The model sample database contains M users, where user m contains N social text data sets p i , p i represents the i-th social text data of the user; the BERT model is used to input the user social text data set and extract features from the user social text data to obtain the text feature h corresponding to the user m i , all text features h of user m i The average pooling process is performed to obtain the initial vector u corresponding to the user; the Llama3 model is used to process the N social text data p of user m. i After the splicing process, the data is extracted and embedded, and then average pooled to obtain the synthesized user integration vector U llama The fusion model includes three fully connected layers and an attention mechanism. The fusion model first transforms the initial vector u and the user integration vector U llama After the conversion, three fully connected layers are used to map the vector to the attention mechanism and generate the vector H llama , then vector H llama Add and fuse with the initial vector u to get the final user vector H of user m p ; The GPT-4 model encodes the personality labels of the model sample database from K dimensions; the target label comparison learning model inputs the user vector H p , personality label encoding for comparative learning model training.

10. The large language model enhanced text mapping processing system for personality assessment according to claim 9, characterized in that: It also includes a new user data collection module, which is used to collect the user's corresponding social text data and process it through the BERT model, L1ama3 model, and fusion model to obtain the user vector H p The target label contrast learning model is based on the input user vector H p Get the final predicted personality label.

Images