Intelligent expert extraction system based on deep learning

Abstract

The invention relates to an intelligent expert extraction system based on deep learning. The intelligent expert extraction system comprises an expert database module, an expert information index module and an intelligent extraction and matching module, the expert database module is used for storing expert information from different fields and carrying out standardization processing; the expert information indexing module is used for intelligently indexing expert information by using a deep learning model and an information retrieval technology, receiving bid evaluation task requirements, screening out experts most conforming to conditions and providing matching degree scores; and the intelligent extraction and matching module converts the bid evaluation task into a task feature vector through an adaptive algorithm, performs similarity calculation on the task feature vector and the expert feature vector, and automatically extracts the most matched expert. According to the method, the standardized expert database is established, intelligent indexing and feature extraction are performed by using the deep learning model, and accurate matching between the experts and the bid evaluation task is realized through the adaptive algorithm, so that high efficiency, justice and transparency of the bid evaluation process are ensured.

Description

Technical Field

[0001] The present invention relates to the field of decision support technology, and in particular to an intelligent expert extraction system based on deep learning. Background Art

[0002] With the continuous advancement of technology and the improvement of industry standards, various bid evaluation activities have gradually transitioned from manual operations to automated and intelligent ones. Traditionally, the selection of experts often relies on manual selection, which is susceptible to human bias and operational errors. The selection process is cumbersome and inefficient, making it difficult to respond to rapidly changing project requirements. Furthermore, the management of expert information is often fragmented and unregulated, leading to inconsistent data and difficulty in querying, further compromising the accuracy and fairness of bid evaluation.

[0003] To improve the fairness and efficiency of the bid evaluation process, in recent years, increasing research and application have focused on the implementation of intelligent expert extraction technology. However, current AI-based expert extraction systems still face several challenges, such as standardized management of expert information, insufficient automation of expert extraction, and optimization issues with extraction algorithms. Therefore, establishing an efficient, intelligent, and impartial expert extraction system has become a key issue in improving the quality and efficiency of bid evaluation.

[0004] While some existing expert extraction methods rely on database queries and rule matching, these methods typically rely on static rules and manual intervention, making them difficult to adapt to complex bid evaluation requirements. Furthermore, when the data volume is large, the system's query efficiency and expert matching accuracy are easily affected. Furthermore, most expert database indexing and information management still rely on traditional technologies, which are unable to meet the growing demand for bid evaluation and lead to bottlenecks in system performance. Summary of the Invention

[0005] In order to solve the above problems, the present invention proposes an intelligent expert extraction system based on deep learning, which can not only improve the efficiency of the bid evaluation process, but also ensure the fairness and transparency of the bid evaluation, providing the industry with a more scientific and efficient solution.

[0006] The specific solution is as follows: An intelligent expert extraction system based on deep learning, including:

[0007] Expert database module, which stores expert information from different fields through the expert database;

[0008] The expert information indexing module uses an expert database indexing algorithm that includes deep learning models and information retrieval technology to intelligently index expert information in the expert database. It receives the requirements of the bid evaluation task, selects the most qualified experts from the expert database, and provides a matching score.

[0009] The intelligent extraction and matching module matches experts with bid evaluation tasks through an adaptive algorithm, converts each task into a task feature vector, and calculates the similarity with the feature vectors of experts screened by the expert information index module, and finally automatically extracts the expert who best matches the task requirements.

[0010] Furthermore, the expert information stored in the expert database is expert information that has undergone expert data normalization processing. The expert data normalization processing is as follows:

[0011] Normalize the collected expert data, remove duplicate expert data and null values, perform format consistency check and cross-domain verification on expert data from different sources, encode and vectorize expert data that meets the standards, and store the converted expert data in the expert database. Automatically connect the expert database with external data sources or platforms, and update the expert database regularly.

[0012] Furthermore, the expert database indexing algorithm comprising a deep learning model and information retrieval technology is used to intelligently index the expert information in the expert database, specifically including:

[0013] Utilize deep learning models to automatically identify and generate indexes, dynamically adjust index strategies based on historical index records, and adaptively optimize indexes based on changes in expert data and query requirements. Through a multi-level index structure, ensure stability in multi-condition queries and high concurrency situations.

[0014] Furthermore, the deep learning model is an EE Transformer model for processing and extracting multi-dimensional expert information; the EE Transformer model processes the sequence information of expert data through a self-attention mechanism, including:

[0015] Calculate the distance between text data and expert data using the following formula:

[0016]

[0017] Among them, σ represents a nonlinear activation function, which is used to introduce sparsity and nonlinear characteristics; α represents an adaptive factor, which is used to balance the static distance of the text sequence and the dynamic changes of the content characteristics; f(d n ,d m ) represents the nonlinear mapping of text content similarity, n and m represent coordinate position information, and support the feature representation extracted by the BERT model; D τ Adaptive adjustment of weights between texts is achieved through σ;

[0018] Calculate the distance of numerical data in expert data as follows:

[0019]

[0020] Among them, λ i Indicates the scaling factor of a specific dimension; |x n,i -x m,i | represents the distance between the features at position n and position m in the i-th dimension; w i represents the weight factor dynamically calculated by the self-attention mechanism, which enables the model to adaptively focus on the more important dimensions of the expert features; D v It means that the sensitivity of the model to numerical features is enhanced by measuring the information difference of different spatial dimensions dim;

[0021] Combined with D τ and D n , the formula for the attention mechanism based on expert data is as follows:

[0022]

[0023] Among them, Q represents query, K represents key, and V represents value; d k Represents the dimension of the key vector, which is used to scale the dot product result to prevent the value from being too large; softmax() means converting the dot product result into a probability distribution; β(Q,K) represents a learnable weight function.

[0024] Furthermore, the influence of text features and numerical features on the extraction results is regulated by weight factors α and β respectively. The calculation formula is as follows:

[0025] EEAttention(Q,K,V)=α×Attention τ (Q,K,V)+β×Attention v (Q,K,V)+γ×

[0026] g(Q,K,V);

[0027] Among them, g(Q,K,V) represents a new feature combination function that captures the complex dependency between text features and numerical features by fusing them; γ represents the parameter used to adjust the weight of g(Q,K,V).

[0028] Furthermore, the expert information indexing module includes a data preprocessing part and a feature extraction part;

[0029] The data preprocessing part: uses the BERT model to convert the expert text data into a vector of fixed length; and performs normalization processing on the expert numerical data, wherein the normalization processing includes the Min-Max normalization method and the Z-score normalization method;

[0030] The Min-Max normalization method is used to linearly map the data to a fixed range so that all features in the expert data have the same dimension. The influence of extreme values ​​is suppressed by introducing logarithmic transformation and weight parameters. The formula is as follows:

[0031]

[0032] Among them, α and β are two controllable weight parameters, whose functions can be set according to task requirements to achieve more flexible feature scaling and increase the discrimination of values; λ∈(0,1) is used to control the decay rate and support adaptive adjustment according to the distribution characteristics of numerical data;

[0033] The Z-score normalization method transforms the expert data into a standard normal distribution through the mean and standard deviation, capturing the local pattern of the expert data. The formula is as follows:

[0034]

[0035] Among them, δ (0<δ<1) represents the parameter that controls the scaling of the logarithmic mapping; u represents the mean of the feature; σ represents the standard deviation of the feature; Z-score represents the normalization operation, which is used to make the input numerical features work at the same scale;

[0036] The feature extraction part adopts a hybrid structure of a multi-layer deep neural network (DNN) combined with a convolutional neural network (CNN) and a recurrent neural network (RNN); wherein the CNN is used to extract local features from the text information of the expert data and uses convolution operations to capture key patterns in the text information;

[0037] The input features are represented as an n-dimensional matrix X, and the convolution kernel is represented as a smaller n-dimensional matrix K. The calculation formula for the output feature map Y is as follows:

[0038]

[0039] Among them, ij,l... represents the position of the output matrix, whose dimension is dim; m, n and l represent the coordinates of the convolution kernel in different dimensions.

[0040] Furthermore, the intelligent extraction and matching module is specifically used to:

[0041] Normalized expert data is obtained through the expert database, and the expert database indexing algorithm IDAlgorithm is used to intelligently index and classify the normalized expert data to obtain an expert dataset that matches the bid evaluation task. The bid evaluation task is converted into a bid evaluation task feature vector and the similarity is calculated with the expert dataset feature vector, and finally the most matching expert is automatically extracted.

[0042] Furthermore, the intelligent extraction and matching module is also used to:

[0043] It provides expert information query, task release and matching result viewing functions; supports users to enter keywords for dynamic screening according to needs, and view the list of experts that meet the requirements and their matching scores; supports real-time monitoring of the dynamics of the bid evaluation process, including the expert extraction process, bid evaluation progress and expert participation, to ensure that all operations and decisions are traceable and the process is transparent and fair; supports automatic notification of extracted experts to participate in bid evaluation activities.

[0044] The present invention adopts the above technical solution and has the following beneficial effects:

[0045] (1) This invention systematically organizes information such as experts' professional background, experience, and evaluation through standardized data collection standards, ensuring the accuracy and consistency of expert data and providing reliable data support for subsequent intelligent extraction and bid evaluation processes;

[0046] (2) The present invention automatically identifies and creates valid indexes by designing an efficient indexing algorithm, thereby reducing query time and improving data retrieval efficiency, ensuring that the system can maintain high performance under high load conditions;

[0047] (3) The present invention uses the intelligent extraction platform IESystem to build a user-friendly interface and real-time monitoring function, ensuring the transparency and fairness of the bid evaluation process, and can automatically notify the extracted experts to participate in the bid evaluation activities, simplifying the operation process and improving the bid evaluation efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 Schematic diagram of the overall architecture of the deep learning-based intelligent expert extraction system according to an embodiment of the present invention;

[0049] Figure 2 This is a data interaction diagram of the deep learning-based intelligent expert extraction system in an embodiment of the present invention. DETAILED DESCRIPTION

[0050] The present invention will be described in further detail below with reference to the embodiments and drawings, but the embodiments of the present invention are not limited thereto.

[0051] like Figure 1 As shown, the present invention provides an intelligent expert extraction system based on deep learning, comprising:

[0052] The expert database module 11 stores expert information from different fields through the expert database.

[0053] Specifically, the expert information stored in the expert database is expert information that has undergone expert data normalization processing, and the expert data normalization processing is as follows:

[0054] Normalize the collected expert data, remove duplicate expert data and null values, perform format consistency check and cross-domain verification on expert data from different sources, encode and vectorize expert data that meets the standards, and store the converted expert data in the expert database. Automatically connect the expert database with external data sources or platforms, and update the expert database regularly.

[0055] Specifically, the information collected to obtain expert data includes name, gender, contact information, affiliation, position, educational background, professional background, professional field, discipline direction, technical expertise, experience and ability in specific fields, work experience and scientific research achievements, and evaluation and historical records;

[0056] Specifically, expert personal information is the basic data of the expert database, including the expert's name, gender, contact information, unit, position and academic background; the expert's professional background includes his or her professional field, subject direction and technical expertise, especially his or her experience and ability in a specific field; this information is crucial to ensure the match between the expert and the bid evaluation project, and helps the system accurately extract experts with relevant professional knowledge; work experience and scientific research results cover the expert's work experience, academic achievements, projects participated in and awards received, etc., which can demonstrate the expert's ability and qualifications, thereby improving the fairness and scientific nature of the bid evaluation process; in addition, expert evaluation and historical records include the expert's performance in the bid evaluation activities, evaluation records and recommendations, etc.; these historical records provide a basis for judging the credibility and reliability of the expert, and provide a reference for the precise screening of the intelligent extraction system; by integrating with external data sources (such as academic databases, industry associations, scientific research institutions, etc.), the expert database can realize the automatic update and supplement of expert information. This approach not only improves the integrity and timeliness of the expert database, but also ensures the diversity and accuracy of expert information.

[0057] Specifically, the collected data should be standardized, encoded, or vectorized to adapt to the corresponding deep learning model. All expert data should be converted into a format suitable for deep learning model processing. Fields such as dates and contact information should be uniformly formatted and converted into numerical or vector form. Information such as professional fields and subject directions should be encoded using standard codes, such as the International Classification of Subjects (ICCS) code, and converted into vector representations that can be input into the model. Removing duplicate data and null values ​​not only improves data quality but also improves training efficiency, avoiding the impact of redundant information on the training effect of the deep learning model. Before input into the model, data integrity and consistency can be ensured through data cleaning and preprocessing. Format consistency checks and cross-domain validation should be performed on data from different sources to ensure that data from different sources are consistent in format and content, avoiding data inconsistencies caused by inconsistent formats or input errors. Expert data should be cleaned and preprocessed, and the corresponding text information should be converted into word vectors or character-level vector representations and normalized to enable efficient model learning. Information in the expert database should be regularly updated through automatic connection with external databases or platforms. This part of information should be converted into a format suitable for model input through automated processing, and the data should be updated in a timely manner to maintain the latest expert background and project experience.

[0058] The expert information indexing module 12 uses an expert database indexing algorithm including a deep learning model and information retrieval technology to intelligently index the expert information in the expert database; receives the requirements of the bid evaluation task, screens out the most qualified experts from the expert database, and provides a matching score.

[0059] Specifically, the expert database indexing algorithm using a deep learning model and information retrieval technology is used to intelligently index the expert information in the expert database, specifically including:

[0060] Utilize deep learning models to automatically identify and generate indexes, dynamically adjust index strategies based on historical index records, and adaptively optimize indexes based on changes in expert data and query requirements. Through a multi-level index structure, ensure stability in multi-condition queries and high concurrency situations.

[0061] Specifically, the deep learning model is an EE Transformer model specifically used for processing and extracting multi-dimensional expert information; the EE Transformer model processes the sequence information of expert data through a self-attention mechanism; wherein, the distance formula used to process text data in expert data is as follows:

[0062]

[0063] Among them, σ represents a nonlinear activation function, which is used to introduce sparsity and nonlinear characteristics; α represents an adaptive factor, which is used to balance the static distance of the text sequence and the dynamic changes of the content characteristics; f(d n ,d m ) represents the nonlinear mapping of text content similarity, n and m represent coordinate position information, and support the feature representation extracted by the BERT model; D τ The weights between texts are adaptively adjusted through σ. The distance formula used to process numerical data in expert data is as follows:

[0064]

[0065] Among them, λ i Indicates the scaling factor of a specific dimension, |x n,i -x m,i | represents the distance between the features at position n and position m in the i-th dimension; w i represents the weight factor dynamically calculated by the self-attention mechanism, which enables the model to adaptively focus on the more important dimensions of the expert features; D v It means that the sensitivity of the model to numerical features is enhanced by measuring the information difference of different spatial dimensions dim;

[0066] Combined with D τ and D n , the formula for the attention mechanism based on expert data is as follows:

[0067]

[0068] Among them, Q represents query, K represents key, and V represents value; d k Represents the dimension of the key vector, which is used to scale the dot product result to prevent the value from being too large; softmax() means converting the dot product result into a probability distribution; β(Q,K) represents a learnable weight function.

[0069] Specifically, the influence of text features and numerical features on the extraction results is regulated by weight factors α and β respectively. The calculation formula is as follows:

[0070] EEAttention(Q,K,V)=α×Attention τ (Q,K,V)+β×Attention v (Q,K,V)+γ×

[0071] g(Q,K,V);

[0072] Among them, g(Q,K,V) represents a new feature combination function that captures the complex dependency between text features and numerical features by fusing them; γ represents the parameter used to adjust the weight of g(Q,K,V).

[0073] Specifically, the expert information indexing module includes a data preprocessing part and a feature extraction part:

[0074] The data preprocessing part: uses the BERT model to convert the expert text data into a vector of fixed length; and performs normalization processing on the expert numerical data, wherein the normalization processing includes the Min-Max normalization method and the Z-score normalization method;

[0075] The Min-Max normalization method is used to linearly map the data to a fixed range so that all features in the expert data have the same dimension. The influence of extreme values ​​is suppressed by introducing logarithmic transformation and weight parameters. The formula is as follows:

[0076]

[0077] Among them, α and β are two controllable weight parameters, whose functions can be set according to task requirements to achieve more flexible feature scaling and increase the discrimination of values; λ∈(0,1) is used to control the decay rate and support adaptive adjustment according to the distribution characteristics of numerical data;

[0078] The Z-score normalization method transforms the expert data into a standard normal distribution through the mean and standard deviation, capturing the local pattern of the expert data. The formula is as follows:

[0079]

[0080] Among them, δ (0<δ<1) represents the parameter that controls the scaling of the logarithmic mapping; u represents the mean of the feature; σ represents the standard deviation of the feature; Z-score represents the normalization operation, which is used to make the input numerical features work at the same scale;

[0081] The feature extraction part adopts a hybrid structure of a multi-layer deep neural network (DNN) combined with a convolutional neural network (CNN) and a recurrent neural network (RNN); wherein the CNN is used to extract local features from the text information of the expert data and uses convolution operations to capture key patterns in the text information;

[0082] The input features are represented as an n-dimensional matrix X, and the convolution kernel is represented as a smaller n-dimensional matrix K. The calculation formula for the output feature map Y is as follows:

[0083]

[0084] Among them, ij,l... represents the position of the output matrix, whose dimension is dim; m, n and l represent the coordinates of the convolution kernel in different dimensions.

[0085] Specifically, CNN-based text feature extraction can help the model identify the professional terms or important symbols of experts in specific fields, thereby improving the accuracy of information indexing; RNN is suitable for experts' time series data (such as the timeline of project participation, the time series of awards received, etc.); RNN can effectively capture temporal dependencies through its hidden state transmission mechanism, so that the historical activity information of experts can be accurately represented and provide more reliable time dimension support for expert matching.

[0086] Specifically, the present invention incorporates a self-attention mechanism based on the Transformer. This mechanism enables the model to focus on the relationships between different parts of the information when processing expert data. In particular, when processing diverse information about experts (such as work experience and scientific research achievements), the self-attention mechanism can effectively weigh the weights of each part, thereby improving the accuracy of information retrieval. Through multi-layer self-attention operations, the model can dynamically adjust the influence of each feature based on the context of the expert information, making the indexing process more accurate. For example, the influence of an expert's scientific research achievements in different fields can be weighted using the self-attention mechanism, thereby accurately recommending suitable experts for specific tasks or projects. In addition, the present invention combines the dual tower model with collaborative filtering. The dual tower model comprises two subnetworks: one for processing expert information (such as personal background and scientific research achievements), and the other for processing task information (such as bid evaluation task requirements, domain-specific needs, etc.). Through the collaborative training of these two subnetworks, the model can generate a matching score between each expert and task, thereby making intelligent recommendations. Collaborative filtering technology, combined with historical expert project participation and other user behavior data, can further improve matching accuracy. By leveraging expert review records and project tags in historical data, the system can infer which experts have performed well in similar tasks, thus supporting matching for future tasks.

[0087] Specifically, the present invention can more flexibly capture the multi-dimensional dependencies in expert data. By introducing adaptive weights and nonlinear processing mechanisms, the EETransformer can efficiently identify and utilize key information during the expert extraction and matching process, thereby improving the overall performance and accuracy of the system.

[0088] The intelligent extraction and matching module 13 matches experts with bid evaluation tasks through an adaptive algorithm, converts each task into a task feature vector, and calculates the similarity with the feature vectors of experts screened by the expert information index module, and finally automatically extracts the expert who best matches the task requirements.

[0089] Specifically, the intelligent extraction and matching module is specifically used to:

[0090] Normalized expert data is obtained through the expert database, and the expert database indexing algorithm IDAlgorithm is used to intelligently index and classify the normalized expert data to obtain an expert dataset that matches the bid evaluation task. The bid evaluation task is converted into a bid evaluation task feature vector and the similarity is calculated with the expert dataset feature vector, and finally the most matching expert is automatically extracted.

[0091] Specifically, the intelligent extraction and matching module is further used to:

[0092] It provides expert information query, task release and matching result viewing functions; supports users to enter keywords for dynamic screening according to needs, and view the list of experts that meet the requirements and their matching scores; supports real-time monitoring of the dynamics of the bid evaluation process, including the expert extraction process, bid evaluation progress and expert participation, to ensure that all operations and decisions are traceable and the process is transparent and fair; supports automatic notification of extracted experts to participate in bid evaluation activities.

[0093] Specifically, the intelligent extraction system IESystem after development includes: 1) The expert database module EPData contains expert information from different fields, including personal information, professional background, work experience, scientific research results, historical bid evaluation records, etc. All expert data are standardized and cleaned to ensure its completeness and accuracy. 2) The expert database indexing algorithm module IDAlgorithm uses deep learning models (such as Transformer, RNN, etc.) and traditional information retrieval technology to intelligently index and classify expert information in the expert database. This module can quickly filter out the most qualified experts from the expert database according to the requirements of the bid evaluation task, and provide a matching score; 3) Intelligent extraction and matching module: This module combines the expert database and indexing algorithm, and matches experts with bid evaluation tasks through an adaptive algorithm. Each task will be converted into a task feature vector, and the similarity with the expert's feature vector will be calculated, and the expert who best matches the task requirements will be automatically extracted.

[0094] Specifically, the developed intelligent extraction system, IESystem, also provides a simple and easy-to-use user interface that supports multiple operations, including expert information query, task posting, and matching result viewing. Specific functions include: 1) Expert Search and Screening: Users can enter relevant keywords based on the requirements of the bid evaluation task, such as professional field, technical expertise, and experience requirements. The platform dynamically filters based on the input criteria and displays a list of experts that meet the requirements. The screening results include the expert's basic information, professional field, historical bid evaluation performance, and matching score. 2) Bid Evaluation Task Posting and Management: Users can post tasks based on the requirements of the bid evaluation project. Upon posting, the system automatically generates feature vectors for the bid evaluation task and matches the most suitable experts based on these feature vectors. Users can perform further screening or confirmation based on the matching results. 3) Intelligent Recommendation and Adjustment: The platform intelligently recommends the most suitable expert combination by analyzing the experts' bid evaluation history, field expertise, and the requirements of the current task in real time. The platform also supports manual intervention, allowing users to adjust the recommendation results according to actual needs. 4) Real-time monitoring and recording of the bid evaluation process: The system can display the dynamics of the bid evaluation process in real time, including the expert selection process, bid evaluation progress, expert participation, etc. All operations and decisions are traceable to ensure the transparency and fairness of the process.

[0095] Specifically, to ensure the fairness of the bid evaluation process, IESystem incorporates several key mechanisms into its design: 1) Multi-dimensional Expert Evaluation System: The platform establishes an expert rating system by comprehensively considering multiple dimensions, including experts' qualifications, historical performance, and technical capabilities. This system is dynamically updated to ensure that each expert's comprehensive evaluation reflects their latest performance and credibility. The system also uses historical expert ratings and participation to avoid overuse of the same expert, ensuring diversity and fairness in the bid evaluation process. 2) Transparency in Task-Expert Matching: All task-expert matching processes are processed using an open and transparent algorithm, and matching results are displayed to users in real time through the platform interface. Users can view each expert's matching score and the system's automatic recommendations to ensure there is no human interference or bias. 3) Debiasing Algorithm: To prevent bias in expert recommendations due to certain factors, the platform incorporates a debiasing mechanism in expert matching. By analyzing historical expert ratings, overlap in domains, and excessive reuse, the system avoids overreliance on a small number of experts, ensuring balanced and diverse task allocation. 4) Expert Feedback and Dynamic Adjustment: After the bid evaluation is complete, the platform allows participating experts and users to provide feedback on the evaluation process and matching results. The system uses this feedback as an important basis for evaluating the performance of the expert matching system, helping to further optimize the expert matching algorithm and evaluation system.

[0096] Specifically, in order to ensure the security and privacy of expert information, IESystem has introduced advanced data encryption technology and multiple security mechanisms. All experts' personal information and bid evaluation records are stored in encrypted form, and identity authentication is required for access to ensure that the data is not accessed by unauthorized personnel. In addition, the system supports audit functions, and all operation records are traceable, ensuring the transparency and security of the data. At the same time, IESystem adopts a modular architecture to support the horizontal expansion of the system and flexible adjustment of functions. With the increase in bid evaluation tasks and the continuous expansion of the expert database, the system can automatically optimize the data processing and matching process to cope with larger-scale data input and task requirements. Through continuous learning and updating, the platform can adapt to the bid evaluation needs of different fields and industries and realize cross-domain intelligent matching.

[0097] Specifically, the main goal of the expert database module 11 is to achieve systematic collection and processing of expert information, so that expert data can be used for the training of deep learning models, thereby ensuring the integrity, accuracy and consistency of expert information and improving the overall operational efficiency of the bid evaluation system. It includes basic data collection and data update, and is responsible for collecting basic expert information, and realizing real-time update of expert data through the integration of external databases (external data sources), thereby ensuring the diversity and timeliness of information in the expert database. In addition, the expert database module 11 is also used to standardize, encode and vectorize the collected data, convert the expert data into a format suitable for deep learning model training, and provide reliable data support for the construction of subsequent algorithms and platforms.

[0098] The expert information indexing module 12 aims to achieve efficient information indexing and intelligent expert extraction, ensuring scientific expert selection and a fair evaluation process. This algorithm utilizes data preprocessing, feature extraction, and self-attention to match experts, enabling collaborative work among modules to optimize the expert data extraction and matching process. Data preprocessing involves preprocessing the text and numerical data of expert information. After cleaning, denoising, and encoding, the text data is encoded and pre-trained using BERT to better capture the contextual semantics of the text and accurately represent the expert's professional background. Numerical data is normalized to unify the data scale, reducing the impact of numerical fluctuations on model training, thereby improving model convergence speed and stability. Data processed by the expert information indexing module is more suitable for input into Transformer-based models, significantly improving the efficiency of subsequent feature extraction and matching. Feature extraction primarily relies on a hybrid CNN and RNN architecture to extract key features such as the expert's professional skills, research achievements, and project experience. CNNs capture local features in expert information, while RNNs, through their temporal dependencies, accurately represent the expert's historical activities, particularly the temporal dimension of their academic and work experience. The combination of these two methods generates a comprehensive feature representation of the expert, supporting subsequent matching. Self-attention and expert matching are at the core of the Transformer, enabling the dynamic weighting of expert information across multidimensional features. Through the self-attention mechanism, the model can flexibly focus on different features and adjust the weighting relationship between them, thereby achieving more efficient expert matching. The self-attention mechanism can also automatically identify and filter irrelevant features, further improving matching accuracy and accurately recommending suitable experts for different types of bid evaluation tasks.

[0099] The intelligent extraction and matching module 13 integrates an expert database and information indexing algorithm, displays results, and enables user interaction, ensuring transparency and fairness in the bid evaluation system. The expert database and information indexing algorithm are tightly integrated to support efficient data access and rapid expert matching. In this module, the information indexing algorithm IDAlgorithm, based on deep learning technology, automatically generates an optimal indexing scheme by analyzing historical query patterns, effectively shortening database retrieval time and improving expert information query efficiency. Expert information in the expert database undergoes strict data standardization to ensure data consistency and accuracy, facilitating expert selection for different tasks in the bid evaluation system. The intelligent extraction and matching module enables real-time information updates and data consistency management, providing efficient and reliable data support. Expert matching and system interaction are the core user interface of the intelligent extraction and matching module, responsible for accurately matching experts with tasks and displaying results. Through the intelligent extraction algorithm, the module can flexibly adopt a variety of expert extraction strategies, including random extraction and recommended extraction, based on the bid evaluation rules to meet different business needs. Based on the extraction results, the system automatically generates notifications, invites matching experts to participate in the bid evaluation, and receives their feedback, ensuring an unmanned process and maintaining confidentiality. Interactive real-time monitoring allows users to monitor the status of various operations during the extraction and bid evaluation process, ensuring transparency and preventing human intervention.

[0100] Specifically, such as Figure 2 As shown, it is a data interaction diagram of the intelligent expert extraction system based on deep learning provided by an embodiment of the present application. Specifically, it includes the expert database EPData, the information indexing method IDAlgorithm, and the specific instance process of intelligent extraction and matching as well as the operation interaction details.

[0101] First, we comprehensively collected basic expert information through various channels to build an initial expert database (EPData). We then connected it to external databases for continuous updates to ensure the comprehensiveness, accuracy, and real-time nature of this information. After data normalization, the expert data was digitized, encoded, and vectorized, transforming it into a format suitable for model training, providing a high-quality data foundation for subsequent analysis.

[0102] The normalized data is normalized using the Min-Max or Z-score method to optimize the data distribution structure and improve model training efficiency. The preprocessed expert data includes text information and time series information: the text data is extracted through local features using CNN, and the time series data is captured through RNN to capture time-dependent characteristics. These feature data are then input into the BERT-based pre-trained Transformer model for further training and fine-tuning to fully explore the deep features of the expert information and achieve intelligent matching and accurate extraction of experts. After feature extraction and model training, the system integrates EPData and IDAlgorithm into the intelligent extraction platform IESystem. Through the platform's embedded modules, the system efficiently displays expert matching results and supports user interaction. The intelligent extraction platform provides real-time monitoring of the entire process to ensure the automation, transparency, and fairness of the bid evaluation system, allowing users to easily screen experts and evaluate bids, improving the efficiency and credibility of the bid evaluation process.

[0103] Although the present invention has been particularly shown and described in conjunction with preferred embodiments, it will be understood by those skilled in the art that various changes in form and details may be made to the present invention without departing from the spirit and scope of the invention as defined in the appended claims, and all such changes are within the scope of protection of the present invention.

Claims

1. An intelligent expert extraction system based on deep learning, characterized in that: include: Expert database module, which stores expert information from different fields through the expert database; The expert information indexing module uses an expert database indexing algorithm that includes a deep learning model and information retrieval technology to intelligently index the expert information in the expert database; receives the requirements of the bid evaluation task, selects the most qualified experts from the expert database, and provides a matching score; The intelligent extraction and matching module matches experts with bid evaluation tasks through an adaptive algorithm, converts each task into a task feature vector, and calculates the similarity with the feature vectors of experts screened by the expert information index module, and finally automatically extracts the expert who best matches the task requirements.

2. The intelligent expert extraction system based on deep learning according to claim 1, characterized in that: The expert information stored in the expert database is the expert information after the expert data normalization process. The expert data normalization process is as follows: Normalize the collected expert data, remove duplicate expert data and null values, perform format consistency check and cross-domain verification on expert data from different sources, encode and vectorize expert data that meets the standards, and store the converted expert data in the expert database. Automatically connect the expert database with external data sources or platforms, and update the expert database regularly.

3. The intelligent expert extraction system based on deep learning according to claim 1, characterized in that: The expert database indexing algorithm including the deep learning model and information retrieval technology is used to intelligently index the expert information in the expert database, specifically including: Use deep learning models to automatically identify and generate indexes, dynamically adjust index strategies based on historical index records, and adaptively optimize indexes based on changes in expert data and query requirements; through a multi-level index structure, ensure stability in multi-condition queries and high concurrency situations.

4. The deep learning-based intelligent expert extraction system according to claim 3, characterized in that: The deep learning model is an EETransformer model for processing and extracting multi-dimensional expert information; the EETransformer model processes the sequence information of expert data through a self-attention mechanism, including: Calculate the distance between text data and expert data using the following formula: Among them, σ represents a nonlinear activation function, which is used to introduce sparsity and nonlinear characteristics; α represents an adaptive factor, which is used to balance the static distance of the text sequence and the dynamic changes of the content characteristics; f(d n ,d m ) represents the nonlinear mapping of text content similarity, n and m represent the coordinate position information, and support the feature representation extracted by the BERT model; D τ Adaptive adjustment of weights between texts is achieved through σ; Calculate the distance of the numerical data in the expert data as follows: Among them, λ i Indicates the scaling factor of a specific dimension; |x n,i -x m,i | represents the distance between the features at position n and position m in the i-th dimension; w i represents the weight factor dynamically calculated by the self-attention mechanism, which enables the model to adaptively focus on the more important dimensions of the expert features; D v It means that the sensitivity of the model to numerical features is enhanced by measuring the information difference of different spatial dimensions dim; Combined with D τ and D n , the formula for obtaining the attention mechanism based on expert data is as follows: Among them, Q represents query, K represents key, and V represents value; k Represents the dimension of the key vector, which is used to scale the dot product result to prevent the value from being too large; softmax() means converting the dot product result into a probability distribution; β(Q, K) represents a learnable weight function.

5. The deep learning-based intelligent expert extraction system according to claim 4, characterized in that: The weight factors α and β are used to adjust the influence of text features and numerical features on the extraction results respectively. The calculation formula is as follows: EEAttention(Q,K,V)=α×Attention τ (Q,K,V)+β×Attention v (Q,K,V)+γ× g(Q, K, V); Among them, g(Q, K, V) represents a new feature combination function, which captures the complex dependency between text features and numerical features by fusing text features and numerical features; γ represents the parameter used to adjust the weight of g(Q, K, V).

6. The deep learning-based intelligent expert extraction system according to claim 5, characterized in that: The expert information indexing module includes a data preprocessing part and a feature extraction part; The data preprocessing part: using the BERT model to convert the expert text data into a vector of fixed length; and performing normalization processing on the expert numerical data, wherein the normalization processing includes a Min-Max normalization method and a Z-score normalization method; The Min-Max normalization method is used to linearly map the data to a fixed range so that all features in the expert data have the same dimension. The influence of extreme values ​​is suppressed by introducing logarithmic transformation and weight parameters. The formula is as follows: Among them, α and β are two controllable weight parameters, whose functions can be set according to task requirements to achieve more flexible feature scaling and increase the discrimination of values; λ∈(0,1) is used to control the decay speed and support adaptive adjustment according to the distribution characteristics of numerical data; The Z-score normalization method transforms the expert data into a standard normal distribution through the mean and standard deviation to capture the local pattern of the expert data. The formula is as follows: Among them, δ (0<δ<1) represents the parameter that controls the scaling of the logarithmic mapping; u represents the mean of the feature; σ represents the standard deviation of the feature; Z-score represents the normalization operation, which is used to make the input numerical features work at the same scale; The feature extraction part: adopts a hybrid structure of a multi-layer deep neural network DNN combined with a convolutional neural network CNN and a recurrent neural network RNN; wherein CNN is used to extract local features from the text information of the expert data, and uses convolution operations to capture key patterns in the text information; The input features are represented as an n-dimensional matrix X, and the convolution kernel is represented as a smaller n-dimensional matrix K. The calculation formula for the output feature map Y is as follows: Among them, ij,l... represents the position of the output matrix, whose dimension is dim; m, n and l represent the coordinates of the convolution kernel in different dimensions.

7. The intelligent expert extraction system based on deep learning according to claim 1, characterized in that: The intelligent extraction and matching module is specifically used for: The normalized expert data is obtained through the expert database, and the normalized expert data is intelligently indexed and classified using the expert database indexing algorithm IDAlgorithm to obtain an expert data set that matches the bid evaluation task. The bid evaluation task is converted into a bid evaluation task feature vector and the similarity is calculated with the expert data set feature vector, and finally the most matching expert is automatically extracted.

8. The deep learning-based intelligent expert extraction system according to claim 1, characterized in that: The intelligent extraction and matching module is also used for: Provides expert information query, task publishing and matching result viewing functions; supports users to enter keywords for dynamic screening according to their needs, and view the list of experts who meet the requirements and their matching scores; It supports real-time monitoring of the dynamics of the bid evaluation process, including the expert selection process, bid evaluation progress and expert participation, ensuring that all operations and decisions are traceable and the process is transparent and fair; it supports automatic notification of selected experts to participate in bid evaluation activities.

Images