Business leader characteristic and management belief analysis method based on long text

Abstract

The invention relates to a commercial leader characteristic and management belief analysis method based on a long text, and relates to the field of natural language processing. According to the method, three deep learning models of GPT, BERT and Bi-LSTM are combined, and firstly, the GPT model is adopted to carry out information extraction and abstract generation on a long text; then, context semantic analysis is carried out through a BERT model, and complex meanings in the text are deeply understood; and finally, carrying out sequence modeling by utilizing a Bi-LSTM model, and capturing a long-distance dependency relationship. The accuracy and generalization ability of the models are improved through effective integration of the three models, ensemble learning and hyper-parameter optimization. According to the invention, the method can achieve the multi-dimensional precise recognition of the leader characteristics and management belief of the commercial leader, and effectively overcomes the limitation of the prior art in the aspects of multi-label classification and long text processing.

Description

Technical Field

[0001] The present invention relates to the field of natural language processing (NLP), and in particular to a method for analyzing leadership traits and management beliefs of business leaders based on long texts. Background Art

[0002] Public statements by business leaders and other public figures not only reflect their management philosophies and strategic thinking but also reveal their inherent leadership traits. This is crucial for understanding organizational decision-making, predicting business development trends, and building public trust. Traditional methods for analyzing leadership traits and beliefs rely primarily on qualitative research and structured analysis techniques, such as word frequency analysis, rule matching analysis, grammatical structure analysis, and content analysis. These methods often require extensive human input, are inefficient, and are susceptible to subjective interference, making them difficult to adapt to the automated analysis needs of large-scale, long text corpora. Furthermore, traditional methods have significant limitations in identifying deep semantics and contextual associations, making it difficult to accurately and comprehensively characterize the multidimensional cognitive characteristics of business leaders.

[0003] With the rapid development of artificial intelligence and big data technologies, deep learning models have achieved significant breakthroughs in natural language processing (NLP), becoming a key technical tool for identifying leadership traits and management beliefs. Pre-trained language models such as BERT (Bidirectional Encoder Representations from Transformers) and GPT (Generative Pre-trained Transformer) possess powerful semantic understanding and context modeling capabilities, and can be used to improve the automation and classification accuracy of text analysis. However, existing models still face input length limitations when processing extremely long texts, often requiring segmentation or summarization. This can weaken the model's grasp of the text's global semantics and affect the recognition of multi-dimensional features. Furthermore, multi-label classification methods combining multi-model architectures remain exploratory in the field of leadership trait identification, and systematic research is lacking on how to improve model accuracy and generalization.

[0004] At present, relevant patents mainly focus on text sentiment analysis and have not yet expanded to the field of multi-label character trait modeling based on ultra-long texts. For example, Chinese patent CN114595693B proposes an online text sentiment analysis method based on the Self-Attention model, but it focuses on emotional tendency judgment and cannot support multi-label output and complex semantic modeling. CN116306678A uses a hierarchical GRU network to process small-scale data sets, which improves the robustness of sentiment analysis, but it is still difficult to meet the needs of extracting multi-dimensional belief structures. CN110889282A introduces a graph convolutional neural network (EmoGCN) to enhance the ability to model lexical relationships, but its scope of application is limited to emotional scenarios and lacks the ability to finely express management styles and strategic tendencies. Summary of the Invention

[0005] In light of this, this paper proposes a method for analyzing the leadership traits and management beliefs of business leaders based on long text. This method addresses the limitations of existing technologies in multi-label classification and long text processing. By combining the generative pre-trained model (GPT), the bidirectional encoder model (BERT), and the bidirectional long short-term memory (Bi-LSTM) network, this method efficiently and accurately identifies and classifies the multi-dimensional leadership traits and management beliefs demonstrated in the public expressions of business leaders.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0007] A method for analyzing leadership traits and management beliefs of business leaders based on long texts, including the following steps:

[0008] Step 1: Build a deep learning network for analyzing leadership traits and management beliefs of business leaders;

[0009] Step 2: Collect public long-text data from business leaders, label their leadership traits and management beliefs using a multi-round review mechanism, and then pre-process the public long-text data from business leaders. The pre-processing includes removing noise information, unifying the format, and performing word segmentation.

[0010] Step 3: Input the pre-processed public long text data of business leaders into the deep learning network in step 1. The deep learning network outputs label predictions of the business leaders' leadership traits and management beliefs. A multi-label binary cross-entropy loss function is used for loss calculation, and the deep learning network is optimized through a backpropagation algorithm. When the loss value converges, the current deep learning network model is saved.

[0011] Step 4: Use the accuracy and harmonic mean F1 score indicators to evaluate the classification effect of the saved deep learning network model. Based on the evaluation results, use the Bayesian optimization method to adjust the hyperparameters to obtain the final deep learning network model.

[0012] Step 5: Preprocess the public long text data of the business leader to be analyzed and input it into the final deep learning network model to obtain the corresponding leadership traits and management beliefs of the business leader to be analyzed.

[0013] Furthermore, the deep learning network in step 1 specifically includes a generative pre-trained language model GPT, a bidirectional encoder pre-trained language model BERT, a bidirectional long short-term memory network Bi-LSTM, a fully connected network, and a classification layer, which are connected in sequence;

[0014] The generative pre-trained language model GPT generates a summary of the pre-processed long text, retaining the main semantic information and reducing redundant content;

[0015] The bidirectional encoder pre-trains the language model BERT for feature extraction and generates high-quality semantic embedding vectors through the bidirectional Transformer structure;

[0016] The bidirectional long short-term memory network Bi-LSTM captures long-distance dependencies in text through its bidirectional structure, enhancing semantic understanding capabilities;

[0017] The fully connected network performs a linear transformation on the output vector of the bidirectional long short-term memory network Bi-LSTM;

[0018] The classification layer uses the Sigmoid activation function to independently predict the leadership traits and management beliefs of each business leader, thereby achieving a multi-label classification task.

[0019] Furthermore, in step 2, public long text data of business leaders including corporate reports, public speeches, interviews and social media posts are collected.

[0020] Furthermore, the deep learning network in step 3 outputs label predictions of business leaders' leadership traits and management beliefs, including: control ability, self-confidence, task orientation, internal trust tendency, external trust tendency, reward and punishment strategies, and risk preference.

[0021] Furthermore, in step 3, regularization technology is introduced during the training of the deep learning network model.

[0022] Furthermore, the final deep learning network model in step 5 provides users with real-time business leader leadership traits and management belief analysis and identification services through the API interface, supports the submission of public long text data of business leaders to be analyzed and returns multi-label classification results.

[0023] Furthermore, the final deep learning network model in step 5 is incrementally trained by periodically updating it with newly collected data.

[0024] Due to the adoption of the above technical solution, the present invention has the following beneficial effects compared with the prior art:

[0025] 1. This paper combines the GPT and BERT models, innovatively addressing the balance between semantic compression and deep understanding of long texts. GPT generates summaries that retain the core information of the text, significantly alleviating the input length bottleneck faced by traditional models; BERT further extracts deep semantic features, providing high-quality input for the subsequent Bi-LSTM architecture.

[0026] 2. The Bi-LSTM method in this invention enhances global information extraction capabilities through bidirectional processing, effectively improving the accuracy of identifying leadership style and management belief dimensions. This overall approach possesses excellent automated processing capabilities and efficient text parsing capabilities, making it suitable for analyzing large-scale commercial text data. It significantly reduces manual intervention costs and improves analytical efficiency and intelligence. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 This is an overall flow chart of a method for analyzing leadership traits and management beliefs of business leaders based on long text in an embodiment of the present invention.

[0028] Figure 2 Schematic diagram of the structure of the BERT model described in an embodiment of the present invention.

[0029] Figure 3 Schematic diagram of the structure of the bidirectional cyclic network layer in an embodiment of the present invention.

[0030] Figure 4 Schematic diagram of the structure of the fully connected layer in an embodiment of the present invention.

[0031] Figure 5 Schematic diagram of the classifier layer structure in an embodiment of the present invention. DETAILED DESCRIPTION

[0032] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0033] A method for analyzing leadership traits and management beliefs of business leaders based on long texts, such as Figure 1 As shown, the overall structure of the system is as follows Figure 1As shown in the figure, it mainly includes a data acquisition module, a data preprocessing module, a summary generation module, a feature extraction module, a deep feature fusion module, a classifier module, and a model training and optimization module. The data acquisition module is responsible for collecting public speech texts of business leaders from multiple countries and regions, including corporate reports, public speeches, interviews, social media, etc. The collected raw text data usually contains a large amount of noise information, such as HTML tags, special characters, and irrelevant advertising content. To ensure the effectiveness of subsequent processing, the data preprocessing module cleans and formats this data, including removing noise information, unifying the text encoding format (such as UTF-8), word segmentation, and text normalization (such as case conversion and lemmatization). The preprocessing process ensures the quality of the input data and lays a good foundation for subsequent model processing.

[0034] After preprocessing, the summary generation module uses the generative pre-trained model GPT to generate key summaries for long texts. Leveraging its powerful text generation capabilities, the GPT model understands the context of the input text and extracts key information from long texts, ensuring that the generated summary is concise while retaining key information, laying a solid foundation for subsequent feature extraction and classification. This step not only addresses the input length limitation of the BERT model when processing extremely long texts, but also preserves the text's key semantic information through summary generation, reducing redundant content and providing strong support for subsequent feature extraction and classification tasks.

[0035] The generated summary text is then input into the feature extraction module, which performs feature extraction based on the BERT model. Figure 2 The figure shows the structure of the BERT model, which demonstrates the multi-layer Transformer architecture of its bidirectional encoder, which can deeply understand the contextual relationships of the text and generate high-quality semantic embedding vectors. The BERT bidirectional encoder pre-trains the language model, which generally accepts a sentence consisting of multiple words as input. Each sentence requires some preprocessing before input, namely adding the [CLS] marker to the beginning of the sentence to indicate the beginning of the sentence and inserting the [SEP] marker to indicate the end of the sentence at the end of the sentence. The input sentence is then converted into three input vector embeddings through preprocessing: word vector, segment vector, and position vector embedding. The final input is the sum of the above three vector embeddings to form a multi-dimensional word embedding vector. After multiple rounds of Transformer transformation, the final encoding vector with text semantics is output.

[0036] In order to further improve the model's ability to capture long-distance dependencies and global semantic information, the embedding vector output by the BERT model is input into the Long Short-Term Memory (LSTM) network. Specifically, the LSTM model is composed of the input word X t , State C t , temporary status Hidden state h t , forget gate f t , Memory Gate i t , output gate O t The calculation process of LSTM is summarized as follows: by forgetting the information in the cell state and memorizing new information, the information useful for subsequent calculations is transmitted, while the useless information is discarded, and the hidden state h is output at each time step. t . The forget gate f t , Memory Gate i t , output gate O t By the hidden state h of the previous moment t-1 and the input X at the current time step t Come to.

[0037] Forget gate: f t =σ(W f ·[h t-1 ,x t ]+b f ),

[0038] Memory Gate: i t =σ(W i ·[h t-1 ,x t ]+b i ),

[0039] Temporary status:

[0040] state:

[0041] Output gate: O t =σ(W O ·[h t-1 ,x t ]+b O )

[0042] Hidden state: h t =O t *tanh(C t )

[0043] σ is the sigmoid activation function, tanh is the tanh activation function

[0044] However, unidirectional LSTM networks can only capture sequence information in one direction and cannot simultaneously model contextual semantic information from both left to right and right to left. Therefore, this paper uses a bidirectional LSTM network, or Bi-LSTM network, to capture global semantic information in both the forward and backward directions, and further extract global semantic features from the time step sequence vector of the hidden state output by the BERT pre-trained language model. Figure 3 The structure diagram of the bidirectional recurrent network layer is shown, and the specific implementation method of the Bi-LSTM network is explained. Through its bidirectional structure, sequence data is processed from both forward and backward directions respectively, capturing long-distance dependencies and global semantic features in the text, and enhancing the ability to understand the text, especially in capturing long-distance dependencies and complex semantic structures.

[0045] Before the feature vector output by the Bi-LSTM network is input to the classifier layer, the present invention uses a fully connected layer as a linear layer to process the high-level global context semantic information representation vector of the sentence into a real-valued vector whose number of dimensions is equal to the total number of categories of the text. The nine-dimensional vector shares the parameter weight w and bias. The structure of the fully connected layer is as follows: Figure 4 As shown in Figure 2, the linear transformation process of the fully connected layer is described, including the specific implementation of the weight matrix and bias term.

[0046] Finally, if Figure 5 As shown, the classifier module performs multi-label classification on the feature vector output by the fully connected layer. The classifier uses the Sigmoid activation function, which can independently predict the probability of each dimension and meet the needs of multi-label classification. Personality classification constructs classification models for the following seven labels: 1. Control over Events; 2. Self-Confidence; 3. Task Orientation; 4. Internal Trust Bias; 5. External Trust Bias; 6. Reward and Punishment Strategy; 7. Risk Propensity. The SoftMax mediation probability value calculation formula of the classifier layer is as follows:

[0047]

[0048] in, represents the predicted probability of the jth label, M is the number of output neurons in the fully connected layer, g jis the jth output value of the fully connected layer. Because each leadership trait dimension is independent, the multi-label classifier can simultaneously predict multiple dimensions, thereby achieving comprehensive identification of business leaders' leadership traits and management beliefs. This process not only improves classification accuracy but also enhances the model's flexibility and adaptability, enabling it to adapt to different business sectors and the changing speaking styles of different corporate leaders.

[0049] During implementation, the model design specifically considered seven key dimensions to identify the multidimensional leadership traits of business leaders: control, self-confidence, task orientation, internal trust, external trust, reward and punishment strategies, and risk appetite. Each label was further subdivided into four levels to accommodate the classification needs of personality traits at different levels. During the data annotation phase, a multi-round review mechanism was implemented to ensure that each business leader's speech text was accurately labeled according to these seven labels.

[0050] The entire model optimizes parameters at each layer through an end-to-end training process to improve classification accuracy and model generalization. During the training process, a multi-label binary cross-entropy loss function (Binary Cross-EntropyLoss) is used to perform independent loss calculations on the prediction results of each label, and the model parameters are optimized through the back-propagation algorithm. The specific loss function formula is as follows:

[0051]

[0052] Among them, Q is the number of samples, θ represents the set of all parameters, is the true value of the jth label of the i-th sample, is the model's predicted probability for the jth label, and λ is the regularization parameter. To prevent overfitting, regularization techniques such as Dropout and L2 regularization are introduced during model training to enhance the model's robustness and generalization. The training process includes data segmentation, model training, model validation, hyperparameter tuning, and model testing. Through supervised learning on large-scale annotated datasets, the model can be continuously iteratively optimized, improving its performance in identifying leadership traits in complex business environments.

[0053] After model training, the model's performance was evaluated on the validation set, using metrics such as accuracy and harmonic mean F1 score (F1 score) to measure the model's classification effectiveness. Based on the evaluation results, model hyperparameters were adjusted or the network structure was optimized to further improve the model's accuracy and generalization. Testing on a dataset of public speeches by business leaders from various fields demonstrated that the model achieved significant results in identifying multidimensional leadership traits and management beliefs. Furthermore, the model was regularly updated and incrementally trained with newly collected data to ensure its timeliness, accuracy, and adaptability. Specific experimental results showed that a model combining GPT, BERT, and Bi-LSTM achieved higher F1 scores and lower misclassification rates in multi-label classification tasks compared to single models or traditional methods, demonstrating the effectiveness and superiority of this approach in practical applications.

[0054] The implementation of the present invention also includes the model deployment and application stage. The trained model is deployed to a cloud server or a local computing environment, and an API interface is provided to support real-time leadership trait recognition requests. The deployment process ensures the efficient operation and scalability of the model. Users can submit long text data through the API, and the model returns the corresponding multi-label classification results of the leadership traits and management beliefs of business leaders, supporting scenarios such as enterprise intelligent decision-making, talent assessment, and market research. System maintenance and iteration include regular updates of model parameters and optimization algorithms, and incremental training or fine-tuning based on newly collected data to ensure the timeliness and adaptability of the model. With the continuous iteration of the model, it has strong cross-domain migration capabilities and can be expanded to multiple high-value application fields such as leadership training, organizational behavior research, public opinion analysis, and brand management in the future. System maintenance also includes monitoring model performance, handling abnormal requests, and ensuring data security to ensure the stability and reliability of the model in actual applications.

[0055] In summary, this paper combines the generative model GPT, the bidirectional encoder model BERT, and the bidirectional long short-term memory network Bi-LSTM to provide an efficient and accurate multi-dimensional method for identifying the leadership traits and management beliefs of business leaders. This method effectively overcomes the limitations of existing technologies in multi-label classification and long text processing, demonstrating significant innovation and broad practical value.

[0056] Those skilled in the art will appreciate that the embodiments described are intended to help readers understand the principles of the present invention and should be understood that the scope of protection of the present invention is not limited to the embodiments described. It will be apparent to those skilled in the art that various modifications and variations are possible in the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention are intended to be included within the scope of the claims of the present invention.

Claims

1. A method for analyzing leadership traits and management beliefs of business leaders based on long texts, characterized by: The following steps are involved: Step 1: Build a deep learning network for analyzing leadership traits and management beliefs of business leaders; Step 2: Collect public long-text data from business leaders, label their leadership traits and management beliefs using a multi-round review mechanism, and then pre-process the public long-text data from business leaders. The pre-processing includes removing noise information, unifying the format, and performing word segmentation. Step 3: Input the pre-processed public long text data of business leaders into the deep learning network in step 1, and the deep learning network outputs label predictions of the leadership traits and management beliefs of business leaders; A multi-label binary cross entropy loss function is used for loss calculation, and the deep learning network is optimized through the back-propagation algorithm. When the loss value converges, the current deep learning network model is saved. Step 4: Use the accuracy and harmonic mean F1 score indicators to evaluate the classification effect of the saved deep learning network model. Based on the evaluation results, use the Bayesian optimization method to adjust the hyperparameters to obtain the final deep learning network model. Step 5: Preprocess the public long text data of the business leader to be analyzed and input it into the final deep learning network model to obtain the corresponding leadership traits and management beliefs of the business leader to be analyzed.

2. The method for analyzing leadership traits and management beliefs of business leaders based on long text according to claim 1, characterized in that: The deep learning network in step 1 specifically includes a sequentially connected generative pre-trained language model GPT, a bidirectional encoder pre-trained language model BERT, a bidirectional long short-term memory network Bi-LSTM, a fully connected network, and a classification layer; The generative pre-trained language model GPT generates a summary of the pre-processed long text, retaining the main semantic information and reducing redundant content; The bidirectional encoder pre-trains the language model BERT for feature extraction and generates high-quality semantic embedding vectors through the bidirectional Transformer structure; The bidirectional long short-term memory network Bi-LSTM captures long-distance dependencies in text through its bidirectional structure, enhancing semantic understanding capabilities; The fully connected network performs a linear transformation on the output vector of the bidirectional long short-term memory network Bi-LSTM; The classification layer uses the Sigmoid activation function to independently predict the leadership traits and management beliefs of each business leader, thereby achieving a multi-label classification task.

3. The method for analyzing leadership traits and management beliefs of business leaders based on long text according to claim 1, characterized in that: In step 2, we collected public long-text data from business leaders, including corporate reports, public speeches, interviews, and social media posts.

4. The method for analyzing leadership traits and management beliefs of business leaders based on long text according to claim 1, characterized in that: The deep learning network described in step 3 outputs label predictions of business leaders' leadership traits and management beliefs, including: control ability, self-confidence, task orientation, internal trust tendency, external trust tendency, reward and punishment strategies, and risk preference.

5. The method for analyzing leadership traits and management beliefs of business leaders based on long text according to claim 1, characterized in that: In step 3, regularization technology is introduced during the training of the deep learning network model.

6. The method for analyzing leadership traits and management beliefs of business leaders based on long text according to claim 1, characterized in that: The final deep learning network model in step 5 provides users with real-time analysis and identification services of business leaders' leadership traits and management beliefs through the API interface. It supports the submission of public long-text data of business leaders to be analyzed and returns multi-label classification results.

7. The method for analyzing leadership traits and management beliefs of business leaders based on long text according to claim 1, characterized in that: The final deep learning network model in step 5 is incrementally trained by periodically updating it with newly collected data.

Images