Personality traits assessment method based on large language model fine tuning

Abstract

The invention discloses a personality traits assessment method based on large language model fine tuning. The method comprises the following steps: firstly, constructing a fine tuning data set, loading a pre-trained large language model, and configuring low-rank adaptation LoRA fine tuning parameters; then large language model training is carried out, the data set training model is finely adjusted based on a classification task to obtain personality dimension scoring levels, classification results serve as auxiliary information, the data set training model is finely adjusted based on a regression task to capture continuity features of personality dimensions, and model performance is improved through iterative optimization; and finally, evaluating the model effect by adopting classification accuracy, mean absolute error, mean square error and Pearson's correlation coefficient indexes. According to the method, the LoRA technology is utilized to perform efficient fine tuning on the large language model, the evaluation accuracy is improved through a strategy of first classification and then regression, the capacity of capturing personality continuity features by the large language model is enhanced, the computing resource demand is reduced, and the method has application value in human resource scenes such as enterprise recruitment and the like.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology, specifically to the application of artificial intelligence technology in human resource management, and relates to a method for evaluating personality traits using a fine-tuned large language model. Background Art

[0002] Personality trait assessment, which identifies an individual's inherent qualities by analyzing their behavior and expressions in specific situations, is of great value in areas such as corporate recruitment and psychological assessment. In the context of asynchronous video interviews, where efficiency is increasingly important, accurately assessing a candidate's personality traits (such as those defined by the HEXACO model) is crucial for predicting their job fit and future performance.

[0003] Existing methods often use traditional machine learning or deep learning models (such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs)) to automatically assess personality traits. However, these methods often rely heavily on large, carefully annotated training datasets. Not only are data acquisition and annotation costs high, but the models are often trained for specific scenarios, resulting in insufficient generalization and difficulty adapting to diverse interview questions, candidate groups, and changing assessment environments. In recent years, large language models (LLMs), with their powerful language understanding and generation capabilities, have made significant progress in natural language processing tasks, providing a new opportunity to address these challenges. Researchers have begun exploring the use of LLMs for personality trait assessment, but current LLM-based methods still have limitations. Most studies focus primarily on classification (e.g., determining personality level) and fail to fully leverage LLMs' ability to process complex information for more sophisticated regression analysis (i.e., predicting specific personality scores), limiting the accuracy and depth of assessments. Furthermore, fine-tuning all parameters of LLMs to adapt to specific tasks often requires significant computational resources and is prone to overfitting. Directly using general-purpose LLMs also struggles to achieve optimal performance in specific scenarios.

[0004] Therefore, there is an urgent need for a new technical solution that can effectively leverage the advantages of large language models and, through more efficient fine-tuning strategies, achieve accurate, fine-grained, and robust assessment of candidate personality traits in asynchronous video interview scenarios while overcoming the limitations of existing methods. Summary of the Invention

[0005] The purpose of this invention is to provide a personality trait assessment method based on fine-tuning a large language model, addressing the technical issues of existing personality trait assessment methods, such as insufficient accuracy, weak generalization, and inadequate analysis of the continuous characteristics of personality traits. Specifically, existing methods rely on traditional models with high annotation costs and limited generalization, or fail to fully utilize large language models for refined regression analysis. Furthermore, fine-tuning all parameters faces the risk of enormous computational resources and overfitting.

[0006] The core of our method lies in employing a two-step evaluation strategy, combined with Low-Rank Adaptation (LoRA) fine-tuning technology, to efficiently and specifically optimize a pre-trained large language model, improving the accuracy and granularity of personality trait assessment in asynchronous video interview scenarios. LoRA fine-tuning is used to fine-tune large language models. When combined with a large language model, it enables specific control over generated results through training with a small amount of data.

[0007] The method of the present invention is specifically as follows: Step (1) Model configuration and construction of fine-tuning dataset: (1-1) Model configuration: First, load the pre-trained open-source causal language model onto the server’s GPU; then create a LoRA fine-tuning configuration for the causal language model, including determining the target module and setting key hyperparameters. The key hyperparameters include the rank of the low-rank matrix. , scaling factor and Dropout ratio; The LoRA fine-tuning technique described is to freeze the original parameter matrix of the causal language model , represents the field of real numbers, Represents the number of rows and columns of the square matrix; introduces two low-rank matrices and , , the original parameter matrix Update to the parameter matrix to be optimized , the number of parameters that need to be optimized in the causal language model is given by Reduce to ; During the model training phase, only two low-rank matrices A and B are optimized, while the original parameter matrix Although it participates in forward and backward propagation calculations, its value remains unchanged (1-2) Construct a fine-tuning dataset: Construct the dataset required for fine-tuning based on the personality trait dataset. Randomly extract some samples from the original dataset to form a fine-tuning dataset, and normalize each data into an instruction-driven triple structure: instruction, input content input and target output output; the input content input includes interview questions and candidate answer text; the target output output is different in different task stages: in the classification task, the target output output is the preset personality dimension score level; in the regression task, the target output output is the specific personality dimension score value.

[0008] Step (2) uses the fine-tuning dataset to train the causal language model to obtain a personality trait assessment model; specifically: (2-1) Fine-tune the classification task of the causal language model: First, design a data preprocessing function, which concatenates the instruction and input content into a conversational prompt that conforms to the model training format, and adds a model-specific end-of-sequence token after the target output output; then input the fine-tuning dataset with the target output output as the personality dimension score level into the causal language model configured with the LoRA fine-tuning technology for fine-tuning, and set the hyperparameters of the training process, including batch size, gradient accumulation steps, learning rate, and number of training epochs; initialize the trainer object and perform model training; after the training is completed, save the learned LoRA weights, merge the LoRA weights with the pre-training causal language model, and generate a classification task optimization model specifically for the personality trait classification task.

[0009] (2-2) Classification Task Reasoning: Design a prompt word template for the reasoning phase of the classification task. The prompt word template integrates the input content and clear instructions for performing the classification task; load the prompt word template into the classification task optimization model, obtain the personality dimension score level predicted by the model, and store it for use in subsequent regression tasks.

[0010] (2-3) Fine-tune the regression task of the causal language model: First, design a data preprocessing function, which concatenates the instruction and input content into a conversational prompt that conforms to the model training format, and adds a model-specific end-of-sequence token after the target output output; then input the fine-tuning dataset with the target output output as the personality dimension score value into the causal language model configured with the LoRA fine-tuning technology for fine-tuning, and set the hyperparameters of the training process, including batch size, gradient accumulation steps, learning rate, and training epochs; initialize the trainer object and perform model training; after training is completed, save the learned LoRA weights, merge the LoRA weights with the pre-training causal language model, and generate a regression task optimization model specifically for personality trait regression tasks.

[0011] (2-4) Regression Task Reasoning: A prompt word template is designed for the reasoning phase of the regression task. The prompt word template integrates the input content, clear instructions for executing the regression task, and the personality dimension score level, aiming to enhance the model's understanding and prediction ability of the continuous numerical characteristics of the personality dimension. The prompt word template is loaded into the regression task optimization model to obtain the specific personality dimension score value predicted by the model.

[0012] (2-5) Iterative optimization: Evaluate the performance of the current model on classification and regression tasks based on performance evaluation indicators. If the performance evaluation indicators do not reach the set values, repeat steps (2-1) to (2-4), and adjust the key hyperparameters set in the LoRA fine-tuning technology configuration or optimize the fine-tuning dataset in this process until the model performance meets the requirements and the final personality trait assessment model is obtained; the optimization of the fine-tuning dataset includes adjusting the data augmentation strategy or the ratio of extracted samples.

[0013] The performance evaluation index includes a classification stage performance evaluation index and a regression stage performance evaluation index.

[0014] The performance evaluation index of the classification stage is the classification accuracy ACC. , is the number of samples for which the model correctly predicts the rating level, is the total number of samples. The higher the classification accuracy, the stronger the ability of the classification task optimization model in preliminarily judging the level of each personality dimension of the candidate.

[0015] The performance evaluation indicators in the regression stage use three indicators: mean absolute error (MAE), mean squared error (MSE), and Pearson correlation coefficient (R) to comprehensively evaluate the accuracy and consistency of the predicted score values of the regression task optimization model.

[0016] , Indicates the The true value of the sample, Indicates the The smaller the MAE value, the smaller the average deviation between the score value predicted by the regression task optimization model and the actual score, and the more accurate the quantitative evaluation.

[0017] ; The lower the MSE value, the smaller the overall error level of the regression task optimization model prediction.

[0018] , and They are the average of the predicted value and the true value respectively; the Pearson correlation coefficient R measures the sequence of predicted values With the true value sequence The degree of linear correlation between ,The closer to 1 the R value is, the stronger the linear correlation between the prediction results of the regression task optimization model and the true labels such as expert reviews, and the more consistent the prediction trend is with the actual situation.

[0019] Step (3) comprehensively evaluates the personality traits of the candidate using the classification task optimization model and the regression task optimization model obtained by the final optimization; specifically: first, the classification task prompt word template is input into the classification task optimization model to obtain the model's predicted candidate personality trait score level result, and then the result is integrated into the regression task prompt word template, and the final regression task prompt word template is input into the regression task optimization model to obtain the model's predicted final candidate personality trait score; finally, the candidate's personality traits are comprehensively evaluated based on the obtained candidate personality trait score.

[0020] By comprehensively analyzing the above classification and regression evaluation indicators, the accuracy, robustness and overall performance of the model trained by the method of the present invention in the personality trait assessment task can be comprehensively and quantitatively evaluated.

[0021] Compared with the prior art, the present invention has the following significant beneficial effects: 1. Improving assessment accuracy and granularity: This paper adopts a two-step strategy of "classification first, then regression", using preliminary classification results as auxiliary information to guide the regression task. This significantly enhances the model's ability to capture the continuous numerical characteristics of personality traits, improves the accuracy and precision of the final score prediction, and overcomes the shortcomings of traditional methods in regression analysis.

[0022] 2. Improve fine-tuning efficiency and reduce resource requirements: Using LoRa fine-tuning technology, only a small number of incremental parameters need to be trained and stored (more than 90% less than full parameter fine-tuning), greatly reducing the demand for computing resources (such as GPU memory) and storage space. At the same time, it effectively alleviates the overfitting problem that is prone to occur when fine-tuning the full parameters of LLMs. This enables LLMs to be adapted and optimized for specific tasks even with limited resources.

[0023] 3. Enhanced generalization and adaptability: Fine-tuning the large language model inherits its rich world knowledge and language understanding capabilities, enabling the model to better understand candidate responses with different expression styles and provide better adaptability and generalization performance to diverse interview questions and candidate backgrounds.

[0024] 4. Filling research gaps and expanding applications: This paper introduces a two-step strategy combined with LoRA fine-tuning in the field of personality trait assessment. In particular, the design of regression tasks fills the gap in existing research and provides new ideas and effective ways for the application of large language models in complex and fine-grained assessment tasks.

[0025] 5. Practical Application Value: This method achieves highly accurate and efficient personality trait assessment at a low computational cost. It is particularly suitable for human resource management scenarios such as corporate recruitment and talent selection. It helps companies more accurately and efficiently screen candidates that are highly matched to job requirements, thereby improving talent competitiveness and organizational effectiveness. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 This is a schematic diagram of the overall framework of the method. It clearly illustrates the complete technical process, from model configuration and fine-tuning dataset construction, to two-step model training (classification task fine-tuning and inference, regression task fine-tuning and inference, and iterative optimization), to final model performance evaluation. DETAILED DESCRIPTION

[0027] In order to make the purpose, technical solutions and advantages of the present invention clearer, the following Figure 1 The specific embodiments of the present invention are described in detail. It should be noted that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0028] like Figure 1, the personality trait assessment method based on large language model fine-tuning, its specific implementation steps are as follows: Step (1) Model configuration and construction of fine-tuning dataset: (1-1) Model configuration: Use Gemma2-7B as the base pre-trained causal language model and load it onto the server's GPU. Create a LoRA configuration instance and specify the target linear layer to which LoRA is applied (for example, the query q_proj, key k_proj, value v_proj, and output o_proj linear layers in the attention mechanism, the up_proj, gate_proj, down_proj linear layers in the feedforward network, etc. The specific selection can be adjusted based on the experiment). Set the core parameters of LoRA: rank (e.g., set to 8 or 16), the scaling factor (usually set to , such as 16 or 32), and the dropout ratio (for example, set to 0.1). Initialize the LoRA configuration and adapt it to the Gemma2-7B model. During subsequent training, ensure that the parameters of the original Gemma2-7B model are frozen, and only the parameters of the low-rank matrices A and B introduced by LoRA are trainable.

[0029] (1-2) Construct a fine-tuning dataset: This dataset is constructed based on the opva database. The database contains the answers of N candidates to 8 questions designed to stimulate specific personality dimensions, as well as the personality dimension scores assessed by experts (including levels of high / medium / low and specific values of 1-5 points). Randomly extract some samples from the original dataset to form a fine-tuning dataset, and organize these data into a unified format. The target output is different in different task stages: in the classification task, the target output is a fine-tuning dataset of personality dimension scores. Each data contains: instruction (for example, "Please judge the performance level of the candidate on the extraversion dimension (high / medium / low) based on the following interview questions and the candidate's answers"), input (including specific question text and candidate answer text), and output (corresponding expert assessment level, such as "high"). In the regression task, the target output is a fine-tuning dataset of personality dimension rating values. Each data entry contains: instruction (for example, "Please predict the candidate's specific score on the extraversion dimension (1-5 points) based on the following interview questions and the candidate's answers"), input (same as above), and output (the corresponding expert evaluation value, such as "4.2").

[0030] Step (2) uses the fine-tuning dataset to train the causal language model to obtain a personality trait assessment model; specifically: (2-1) Fine-tune the classification task of the causal language model: Data preprocessing: Define a function to convert each record in the classification task fine-tuning dataset into the input format accepted by the model. For example, formatted as: <s> User:"+instruction+"\n Input:"+input+"\nAssistant:"+output+"< / s> ",in <s> and< / s> are model-specific start and end markers.

[0031] LoRA fine-tuning: Use the Trainer class in the Hugging Face Transformers library or a similar framework for training. Configure training parameters: batch size (e.g., 4 or 8), number of gradient accumulation steps (e.g., 2 or 4 to increase the effective batch size), learning rate (e.g., 2×10 −4 or 5×10 −5 ), number of training rounds (for example, 3 or 5), logging frequency (logging_steps), model saving frequency (save_steps), etc. Pass the preprocessed classification dataset and the Gemma2-7B model configured with LoRA into Trainer and call the train() method to start the fine-tuning process.

[0032] Model merging and saving: After training, the Trainer automatically saves the optimal LoRA weights. Subsequently, the corresponding function is called to merge the trained LoRA weights with the original Gemma2-7B base model to generate a model optimized for the personality trait classification task. This merged classification task optimized model is saved for subsequent inference.

[0033] (2-2) Classification task reasoning: Load the classification task optimization model saved in step (2-1). For each candidate answer sample that needs to be evaluated, construct a prompt word template for classification task reasoning. The format is similar to the input part during training, but does not contain the target output. For example: <s>User: Based on the following interview questions and the candidate's answers, please judge their performance level on the Extraversion dimension (high / medium / low).\n Input: Question: [Question text]\n Answer: [Candidate's answer text]\n Assistant: ", input the prompt word into the classification task optimization model and use the classification task optimization model's generate() method to obtain the prediction result. Parse the generated text, extract the predicted rating level (such as "medium"), and store it for use in subsequent regression tasks.

[0034] (2-3) Fine-tuning the causal language model for regression: Using a data preprocessing pipeline and LoRA fine-tuning process similar to step (2-1), the fine-tuning dataset, whose target output is the numerical value of the personality dimension score, is fed into the Gemma2-7B model equipped with LoRA for training. After training, the learned LoRA weights are saved and merged with the original Gemma2-7B base model to generate a model optimized specifically for personality trait regression tasks.

[0035] (2-4) Regression Task Reasoning: Load the regression task optimization model saved in step (2-3) and construct a prompt word template for reasoning for each candidate's answer sample to be evaluated. The key point is that the prompt word includes the personality dimension score prediction result obtained in step (2-2) for the sample. For example: <s>User: Based on the following interview questions, the candidate's answers, and their estimated extraversion rating ([classification result]), please predict their specific score on the extraversion dimension (1-5 points).\nInput: Question: [Question text]\nAnswer: [Candidate answer text]\nEstimated rating: [classification result]\nAssistant: "(Note: [classification result] here should be replaced with the actual predicted rating obtained in step (2-2), such as "medium"). Input the prompt word that integrates the predicted personality dimension rating results into the regression task optimization model to obtain the generated text. Finally, parse the generated text and extract the specific predicted rating value (such as "3.8").

[0036] (2-5) Iterative optimization: Collect a batch of test samples (not used for training), use the classification task optimization model and the regression task optimization model for inference, and obtain the prediction results. Calculate the performance of the model on the test set according to the defined evaluation indicators (ACC, MAE, MSE, R). If the performance does not meet the standard (for example, ACC is lower than the preset threshold, or MAE is higher than the preset threshold), return to step (2-1) and adjust the LoRA configuration (for example, increase or decrease the rank ), training hyperparameters (e.g., adjusting the learning rate, increasing the number of training rounds), or improving the fine-tuning dataset (e.g., increasing the amount of data, optimizing data cleaning or augmentation methods), and then re-fine-tune and re-evaluate the original Gemma2-7B model until the model performance meets the requirements.

[0037] Finally, a comprehensive evaluation of the final optimized models for the classification and regression tasks was conducted using an independent test dataset (with no overlap between the training and validation sets). Classification evaluation: The classification task optimization model's accuracy (ACC) in predicting ratings on the test set was calculated. Regression evaluation: The regression task optimization model's MAE, MSE, and R were calculated for predicting specific ratings on the test set. These evaluation metrics were recorded and reported to demonstrate the effectiveness of the proposed method. For example, reporting an ACC above 85%, a MAE below 0.5, and an R above 0.6 indicated that the final personality trait model had good evaluation performance.

[0038] Step (3) comprehensively evaluates the personality traits of the candidate using the classification task optimization model and the regression task optimization model obtained by the final optimization; specifically: first, the classification task prompt word template is input into the classification task optimization model to obtain the model's predicted candidate personality trait score level result, and then the result is integrated into the regression task prompt word template, and the final regression task prompt word template is input into the regression task optimization model to obtain the model's predicted final candidate personality trait score; finally, the candidate's personality traits are comprehensively evaluated based on the obtained candidate personality trait score.

[0039] Through the above steps, this embodiment shows in detail how to use the personality trait assessment method based on large language model fine-tuning proposed in the present invention to achieve accurate and efficient assessment of personality traits.< / s> < / s>

Claims

1. A personality trait assessment method based on fine-tuning of a large language model, characterized by: Step (1) Model configuration and construction of fine-tuning dataset: (1-1) Model configuration: First, load the pre-trained open-source causal language model onto the server’s GPU; then create a LoRA fine-tuning configuration for the causal language model, including determining the target module and setting key hyperparameters. The key hyperparameters include the rank of the low-rank matrix. , scaling factor and Dropout ratio; (1-2) Constructing a fine-tuning dataset: Construct the dataset required for fine-tuning based on the personality trait dataset. Randomly extract some samples from the original dataset to form a fine-tuning dataset, and normalize each data into an instruction-driven triple structure: instruction, input content input, and target output output; the input content input includes interview questions and candidate answer text; the target output output is different in different task stages: in the classification task, the target output output is the preset personality dimension score level; in the regression task, the target output output is the specific personality dimension score value; Step (2) uses the fine-tuning dataset to train the causal language model to obtain a personality trait assessment model; specifically: (2-1) Fine-tune the classification task of the causal language model: First, design a data preprocessing function that concatenates the instruction and input content into a conversational prompt that conforms to the model training format, and adds a model-specific end-of-sequence token after the target output output; then input the fine-tuning dataset with the target output output as the personality dimension score level into the causal language model configured with the LoRA fine-tuning technology for fine-tuning, and set the hyperparameters of the training process, including batch size, gradient accumulation steps, learning rate, and number of training epochs; initialize the trainer object and perform model training; after training is completed, save the learned LoRA weights, merge the LoRA weights with the pre-training causal language model, and generate a classification task optimization model specifically for personality trait classification tasks; (2-2) Classification Task Reasoning: Design a prompt word template for the reasoning phase of the classification task. The prompt word template integrates the input content and clear instructions for performing the classification task; load the prompt word template into the classification task optimization model, obtain the personality dimension score level predicted by the model, and store it for use in subsequent regression tasks; (2-3) Fine-tune the regression task of the causal language model: First, design a data preprocessing function that concatenates the instruction and input content into a conversational prompt that conforms to the model training format, and adds a model-specific end-of-sequence token after the target output output; then input the fine-tuning dataset with the target output output as the personality dimension score value into the causal language model configured with the LoRA fine-tuning technology for fine-tuning, and set the hyperparameters of the training process, including batch size, gradient accumulation steps, learning rate, and training epochs; initialize the trainer object and perform model training; after training is completed, save the learned LoRA weights, merge the LoRA weights with the pre-training causal language model, and generate a regression task optimization model specifically for the personality trait regression task; (2-4) Regression Task Reasoning: Design a prompt word template for the reasoning phase of the regression task. The prompt word template integrates the input content, clear instructions for executing the regression task, and the personality dimension score level, aiming to enhance the model's understanding and prediction ability of the continuous numerical characteristics of the personality dimension. Load the prompt word template into the regression task optimization model to obtain the specific personality dimension score value predicted by the model; (2-5) Iterative optimization: Evaluate the performance of the current model on the classification and regression tasks based on the performance evaluation indicators. If the performance evaluation indicators do not reach the set values, repeat steps (2-1) to (2-4). During this process, adjust the key hyperparameters set in the LoRA fine-tuning technology configuration or optimize the fine-tuning dataset until the model performance meets the requirements, and obtain the final personality trait assessment model; the optimization of the fine-tuning dataset includes adjusting the data augmentation strategy or the ratio of extracted samples; Step (3) comprehensively evaluate the personality traits of the candidate using the finally optimized classification task optimization model and regression task optimization model; Specifically, the classification task prompt word template is first input into the classification task optimization model to obtain the model's predicted candidate personality trait score level result, and then the result is integrated into the regression task prompt word template. The final regression task prompt word template is input into the regression task optimization model to obtain the model's predicted final candidate personality trait score; finally, the candidate's personality traits are comprehensively evaluated based on the obtained candidate personality trait score.

2. The personality trait assessment method based on large language model fine-tuning according to claim 1, characterized in that: The LoRA fine-tuning technique described is to freeze the original parameter matrix of the causal language model , represents the field of real numbers, Represents the number of rows and columns of the square matrix; introduces two low-rank matrices and , , the original parameter matrix Update to the parameter matrix to be optimized , the number of parameters that need to be optimized in the causal language model is given by Reduce to ; During the model training phase, only two low-rank matrices A and B are optimized, while the original parameter matrix Although it participates in forward and backward propagation calculations, its value remains unchanged.

3. The personality trait assessment method based on large language model fine-tuning according to claim 1, characterized in that: The performance evaluation index includes a classification stage performance evaluation index and a regression stage performance evaluation index; The performance evaluation index of the classification stage adopts the classification accuracy ACC, , is the number of samples for which the model correctly predicts the rating level, is the total number of samples. The higher the classification accuracy, the stronger the ability of the classification task optimization model in preliminarily judging the level of each personality dimension of the candidate. The regression stage performance evaluation index uses three indicators: mean absolute error (MAE), mean square error (MSE), and Pearson correlation coefficient (R) to comprehensively evaluate the accuracy and consistency of the predicted score values of the regression task optimization model; , Indicates the The true value of the sample, Indicates the The predicted value of each sample; the smaller the MAE value, the smaller the average deviation between the score value predicted by the regression task optimization model and the actual score, and the more accurate the quantitative evaluation; ； The lower the MSE value, the smaller the overall error level of the regression task optimization model prediction; , and They are the average of the predicted value and the true value respectively; the Pearson correlation coefficient R measures the sequence of predicted values With the true value sequence The degree of linear correlation between ,The closer to 1 the R value is, the stronger the linear correlation between the prediction results of the regression task optimization model and the true labels such as expert reviews, and the more consistent the prediction trend is with the actual situation.

Images