A university AI intelligent one-stop service system based on an agent platform

Abstract

The present invention relates to the field of intelligent service platforms, and discloses a university AI intelligent one-stop service system based on an agent platform, including: an SQL processing module for converting natural language input by a user into SQL statements of a target database; a vector library management module for docking and adapting with various mainstream open-source and closed-source vector libraries and interacting with different vector libraries through a unified API interface; an extensible architecture module that supports multiple models and multiple inference frameworks and processes data from databases, web pages or PDF files; a vertical domain expertise module for performing data preprocessing and pre-training models on university vertical domain data, outputting material completeness scores, missing item detection, and urgency classification information, and adjusting parameter settings according to industry characteristics, and customizing and optimizing the interaction experience through role-based prompt words. This application can intelligently process and analyze multi-source data, significantly improving the business handling efficiency and user experience of the university online service hall.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent service platforms, and specifically to a university AI intelligent one-stop service system based on an agent platform. Background Art

[0002] With the in-depth development of university informatization construction, the online service hall has become an important channel for teachers and students to handle campus business. However, in the actual use process, teachers and students face many challenges. First of all, there are a wide variety of services in the online service hall, and it is difficult to find the entrances. Especially in complex business fields such as finance, the existence of multiple entrances makes it difficult for teachers and students to quickly locate the correct entrance when handling business such as reimbursement. Secondly, teachers and students often do not understand the business rules well, resulting in frequent obstacles due to missing materials during the handling process. This not only wastes time but also affects the smooth handling of business. In addition, the lack of process guidance is also a major problem. Teachers and students need to continuously query the process guidance when handling business, resulting in slow progress in business handling and poor user experience.

[0003] To solve the above problems, universities have begun to explore the introduction of AI technology into the online service hall, and through API docking between the agent platform and the ports of each business system in the online service hall, intelligent business handling is realized. However, the existing agent platforms still have deficiencies in many aspects: the existing platforms lack flexible support for multiple vector databases, and it is difficult to handle the storage and management of large-scale data. Especially in high-concurrency scenarios, the efficiency of data access and processing is low. The natural language processing models used in mainstream GPT platforms are relatively weak in the professionalism of database operations, and it is difficult to accurately understand and generate complex database query statements; it is difficult to support users to flexibly expand and customize the model according to specific needs and different database environments; when dealing with problems in specific industries such as universities, it may not be able to provide in-depth and accurate answers like professionals. Summary of the Invention

[0004] The purpose of the present invention is to solve the problems existing in the prior art, and to propose an AI large model agent platform system applied to the education and government industries, aiming to solve the problems such as difficult entrance finding, lack of understanding of business rules, lack of process guidance, etc. existing in the process of handling business in the university online service hall through intelligent means, as well as the technical problems existing in the existing agent platforms, such as low efficiency of data access and processing, difficulty in accurately understanding and generating complex database query statements; difficulty in supporting users to flexibly expand and customize the model according to specific needs and different database environments; inability to provide in-depth and accurate answers, etc.

[0005] To achieve the above purpose, the present invention adopts the following technical solutions:

[0006] A university AI intelligent one-stop service system based on an agent platform, comprising:

[0007] The SQL processing module is used to organize the SQL statements of existing mainstream databases, and convert the natural language input by users into SQL statements of the target database through pre-trained machine learning or deep learning models;

[0008] The vector library management module is docked and adapted with various mainstream open-source and closed-source vector libraries, interacts with different vector libraries through a unified API interface, and supports dynamic data synchronization and multi-modal joint retrieval;

[0009] The extensible architecture module supports multiple models and multiple inference frameworks, including submitting queries, the retriever component selecting relevant information, and the re-ranker refining the selection, for processing data from databases, web pages or PDF files, and continuously optimizing the data-driven engine through proxy and plugin mechanisms, and ensuring the scalability of the platform through the construction of a knowledge base and intelligent agent workflow components; user-defined data types, indexing strategies and query optimization plugins;

[0010] The vertical domain expertise module preprocesses data and pre-trains models for data in the vertical domain of universities, outputs material completeness scores, missing item detection and urgency classification information, adjusts parameter settings according to industry characteristics, and customizes and optimizes the interaction experience through role-based prompt words.

[0011] Furthermore, the SQL processing module includes:

[0012] The requirement analysis unit is used to determine the target database type, the type of SQL statements to be supported, and the conversion requirements from natural language to SQL;

[0013] The data collection unit is used to collect database SQL statement samples, natural language query samples and their corresponding SQL statements;

[0014] The data preprocessing unit is used to standardize the collected SQL statements and perform natural language processing operations on natural language queries, including word segmentation and part-of-speech tagging;

[0015] The model selection and training unit is used to select machine learning or deep learning models and train the models with the preprocessed data so that they can understand natural language and generate corresponding SQL statements;

[0016] The model optimization unit is used to evaluate the model, optimize the model according to the evaluation results, and integrate the trained model into the intelligent agent platform.

[0017] Furthermore, the vector library management module includes:

[0018] The document acquisition unit is used to obtain the API interface documents of each vector library from the preselected list of vector libraries;

[0019] A platform API interface unit for providing unified API interfaces, enabling the platform to interact with different vector libraries through these interfaces, and the abstraction layer of the API interfaces can handle the differences between different vector libraries.

[0020] Furthermore, the extensible architecture module includes:

[0021] A plugin architecture unit for integrating third-party developers' custom plugins and providing necessary tools and libraries;

[0022] A data type and index specification definition unit for defining the specifications of new data types and customized indexes; wherein, the specifications of the new data types include data structures and operation interfaces, and the specifications of the customized indexes include index structures and index methods;

[0023] A query optimization unit for implementing a query optimizer and supporting user-defined query optimization strategies;

[0024] A plugin management tool for facilitating users to install, configure, and uninstall plugins;

[0025] A testing and verification unit for testing the plugin framework and new functions, and verifying whether the new data types and index types meet the data storage requirements of specific domains.

[0026] Furthermore, the vertical domain expertise module includes:

[0027] A data collection and preprocessing unit for collecting university industry-specific datasets including text and voice data, and performing data preprocessing, including data cleaning, annotation, and structuring;

[0028] A model selection and training unit for processing materials through a multi-task learning framework, selecting appropriate machine learning or deep learning algorithms, training the model using the preprocessed industry-specific datasets, and adjusting parameters according to the model performance to achieve multi-modal material pre-review;

[0029] A role division unit for customizing different prompt words and interaction logics according to the user organization structure roles, and simulating the tone, intonation, and speech rate of university industry users;

[0030] A user feedback unit for collecting user feedback and performing iterative optimization based on the feedback.

[0031] Furthermore, the SQL processing module further includes an agent interaction layer. The agent interaction layer uses a Bi-LSTM+CRF hybrid model to implement business intent recognition, constructs a dynamic business topology map, dynamically adjusts node weights according to real-time access volume, provides recommendation entries and confidence scores for users, and supports multi-turn dialogue management, and can perform context understanding based on user feedback.

[0032] Furthermore, the SQL processing module further includes an NLP engine, a rule engine, and a speech engine, which are respectively used for intent recognition, rule inference, and speech processing;

[0033] Among them, the NLP engine integrates intent recognition and multi-turn dialogue management. The model selection and training unit trains SQL statements through the machine learning or deep learning algorithm to realize the conversion from natural language to SQL statements, and supports college users to perform data query operations;

[0034] The rule engine dynamically loads SWRL rules provided by each business department and supports the automated management of business processes;

[0035] The speech engine supports mixed Chinese and English speech processing, uses the Conformer model for speech recognition, and is fine-tuned on college scenario corpora, and at the same time provides multiple speech synthesis modes.

[0036] Furthermore, the Bi-LSTM+CRF hybrid model is constructed by combining a bidirectional long short-term memory network and a conditional random field, including:

[0037] An input layer, which is used to receive text data and convert each character or word into a word vector representation;

[0038] A word embedding layer, which is used to map word vectors to a high-dimensional space to capture semantic relationships between words;

[0039] The Bi-LSTM layer consists of a forward LSTM and a backward LSTM, which can capture both forward and backward information in the sequence and the output of each time step contains the context information of the current position;

[0040] A fully connected layer, which is used to map the output of the Bi-LSTM layer to the label space and output the score of each position corresponding to each label;

[0041] The CRF layer is used to globally optimize the output of the fully connected layer according to the dependency relationship between labels, and outputs the optimal label sequence by calculating the transition matrix and the emission matrix.

[0042] Furthermore, the SQL processing module further includes a data service layer, which is connected to multiple business databases in real time, constructs a cross-system data federation query interface, and integrates a multi-modal feature extractor; the data service layer is connected to multiple business databases of teaching affairs, finance, and student affairs in real time, constructs a cross-system data federation query interface, and supports the storage, processing, and analysis of large-scale data. By integrating a multi-modal feature extractor, it can extract and analyze features of various types of data such as text, images, and tables, providing rich data support for upper-layer applications.

[0043] Furthermore, the vertical domain expertise module further includes a natural language processing model and a speech recognition model, which are respectively used to process text and voice data;

[0044] The deep learning algorithm of the speech recognition model includes a Transformer model for processing long sequence data and capturing global dependencies, and provides a rich industry knowledge base and a question-and-answer system, supporting users to query industry knowledge and get answers to questions through natural language.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] (1) By enhancing the intelligent agent interaction layer, the present invention realizes more accurate business intention recognition, provides more accurate business entrance recommendations for teachers and students, and greatly shortens the time to find business entrances; by integrating the NLP engine and the rule engine, the conversion from natural language to SQL statements is realized, supporting teachers and students to directly query the database through voice or text without manually writing SQL statements, improving the query efficiency.

[0047] (2) By constructing and real-time adjusting the dynamic business topology graph, the system of the present invention can dynamically adjust the node weights according to the real-time access volume, provide more reasonable business entrance sorting and confidence scores for users, improve the user experience, and the multi-round dialogue management function supports complex dialogue processes, can understand the context according to user feedback, and realizes a more natural and smooth interaction experience.

[0048] (3) The vector library management module of the present invention supports a variety of mainstream open-source and closed-source vector libraries, realizes dynamic data synchronization and multi-modal joint retrieval, and improves the flexibility and efficiency of data processing. The data service layer is connected to multiple business databases in real time, constructs a cross-system data federation query interface, supports the storage, processing, and analysis of large-scale data, and provides rich data support for upper-layer applications.

[0049] (4) The extensible architecture module of the present invention supports multiple models and multiple inference frameworks. Users can flexibly expand and customize the models according to specific requirements and different database environments, improving the scalability of the system. The vertical domain expertise module has performed a large amount of data preprocessing and pre-training for specific industries such as universities, improving the system's ability to answer questions in specific industries. Through role-based prompt customization and a rich industry knowledge base, the system can simulate the tone, intonation, and speech rate of university industry users, providing an interaction experience closer to user needs.

[0050] In summary, by integrating SQL processing capabilities, vector library management, extensible architecture, and vertical domain expertise applications, the present invention realizes the intelligent processing and analysis of multi-source data, significantly improves the business handling efficiency and user experience of the university online service hall, and contributes to the digital transformation and intelligent upgrading of the education industry. Brief Description of the Drawings

[0051] Figure 1 It is a schematic diagram of the modules of a university AI intelligent one-stop service system based on an agent platform proposed in an embodiment of the present invention.

[0052] Figure 2 It is the overall platform architecture of a university AI intelligent one-stop service system based on an agent platform proposed in an embodiment of the present invention;

[0053] Figure 3 It is a schematic diagram of the construction process of the SQL processing module of an AI large model agent platform system applied to the education and government industries proposed in an embodiment of the present invention.

[0054] Figure 4 It is a schematic diagram of the construction process of the vector library management module of a university AI intelligent one-stop service system based on an agent platform proposed in an embodiment of the present invention.

[0055] Figure 5 It is a schematic diagram of the construction process of the extensible architecture module of a university AI intelligent one-stop service system based on an agent platform proposed in an embodiment of the present invention.

[0056] Figure 6 It is a schematic diagram of the construction process of the vertical domain expertise module of a university AI intelligent one-stop service system based on an agent platform proposed in an embodiment of the present invention. Detailed Embodiments

[0057] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0058] In the description of the present invention, it should be noted that the orientation or positional relationship indicated by the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc. is based on the orientation or positional relationship shown in the accompanying drawings. It is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as a limitation to the present invention. In addition, the terms "first", "second", and "third" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance.

[0059] In addition, the technical features involved in different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0060] Embodiment 1

[0061] Refer to Figures 1-6 As shown, a university AI intelligent one-stop service system 100 based on an agent platform includes:

[0062] An SQL processing module 110, which is used to sort out the SQL of existing mainstream databases, and convert the natural language input by users into SQL statements of the target database through a pre-trained machine learning or deep learning model;

[0063] Among them, the target database type (such as MySQL, PostgreSQL, etc.) stores data such as student information, course information, and library collections of the school. The supported SQL statement types (such as SELECT, WHERE, etc.) can meet the data query needs of teachers and students. Analyze the conversion requirements from natural language to SQL, including the complexity of queries (such as multi-table joint queries) and flexibility (such as support for fuzzy queries), etc.

[0064] A vector library management module 120, which docks and adapts with various mainstream open-source and closed-source vector libraries, interacts with different vector libraries through a unified API interface, and supports dynamic data synchronization and multi-modal joint retrieval;

[0065] Among them, the mainstream open-source and closed-source vector libraries include Milvus, Pinecone, Vespa, etc. According to the characteristics, performance, and API interfaces of each vector library, the most suitable vector library for the school's needs is selected. Through a unified API interface, the platform can interact with different vector libraries through these interfaces, ensuring that the abstraction layer of the API interface can handle the differences between different vector libraries and realizing the transparent access of the vector library. For example, in scientific research project management, the system can automatically extract key information (such as project name, person in charge, research content, etc.) from project documents and store it in the vector library. When teachers and students need to query information related to a certain project, the system can quickly find relevant text, images and other data through multimodal joint retrieval and display them to users.

[0066] The extensible architecture module 130 supports multiple models and multiple inference frameworks, including submitting queries, the retriever component selecting relevant information, and the reranker refining the selection, for processing data from databases, web pages, or PDF files, and continuously optimizing the data-driven engine through proxy and plugin mechanisms, and ensuring the scalability of the platform through the construction of the knowledge base and the intelligent agent workflow component; user-defined data types, indexing strategies, and query optimization plugins;

[0067] Among them, through the proxy mechanism, transparent access and scheduling of the data-driven engine are realized, and through the plugin mechanism, the performance and functions of the data-driven engine are continuously optimized to adapt to new business requirements and data sources. For example, in the library management system, the system supports the storage and retrieval of multiple data types, such as book information, borrowing records, etc., allowing users to customize data types (such as text, images, audio, etc.), indexing types (such as full-text indexing, vector indexing, etc.), and query optimization strategies (such as caching strategies, query rewriting, etc. data types) through the graphical interface or API interface, automatically creating corresponding data structures according to user definitions, optimizing the storage and retrieval efficiency of book information, and the system integrates query optimization plugins, such as caching plugins, to improve the query speed of borrowing records.

[0068] The vertical domain expertise module 140 performs a large amount of data preprocessing and pre-trained models (such as multi-task learning frameworks) on data in the vertical domain of universities, outputs material completeness scores, missing item detection, and urgency classification information, and adjusts parameter settings according to industry characteristics, and optimizes the interaction experience through role-based prompt word customization.

[0069] Among them, the system outputs material completeness scores, missing item detection, and urgency classification information. For example, the system can check the integrity, missingness, and urgency of the materials submitted by the user (such as reimbursement vouchers, application forms, etc.), and give the material completeness scores respectively. This score can be a specific value (such as 0-100 points) or a level (such as "complete", "basically complete", "missing some materials", etc.). According to the score, it is decided whether to continue processing or prompt the user to supplement the materials; identify the missing materials or information items, inform the user of what materials or information are missing, help the user quickly complete the materials, and improve the business handling efficiency; classify the urgency of the business according to factors such as the business type submitted by the user and the current time (such as "urgent", "general", "non-urgent", etc.), help the system give priority to processing urgent business, and improve the business processing efficiency. At the same time, it also provides reference information for business handling for users, and users can reasonably arrange the handling time according to the urgency.

[0070] Among them, the vertical domain expertise module 140 performs a large amount of data preprocessing on data in the university vertical domain. For example, it collects university industry-specific data sets, including relevant text and voice data for business such as scholarship applications and reimbursements, and performs data cleaning, annotation, and structuring for subsequent model training. Role-based prompt customization, for example, is divided according to the user's organizational structure role (such as students, teachers, administrative staff, etc.), and different prompts and interaction logics are customized. For example, for student users, the system can use more friendly and easy-to-understand prompts and interaction methods.

[0071] Specifically, such as Figure 2As shown in the figure, the overall platform architecture of the university AI intelligent one-stop service system 100 based on the agent platform includes 7 levels: (1) Basic environment level: The platform supports industry-specific cloud environments such as campus networks and educational affairs clouds, and also supports private deployment and mainstream public cloud environments such as Alibaba Cloud. (2) Model layer: Currently, the platform builds a dedicated large model for text-to-SQL for SQL statements and various industry-specific models for universities, governments, etc. At the same time, the platform supports various current mainstream large models and various third-party open-source and closed-source large models such as Tongyi Qianwen, Doubao, and Kimi. (3) Protocol layer: The platform supports various current mainstream models and interface database interfaces, including DSL, AgentFrame, Operators, Parallelization, etc. (4) Module layer: The platform includes a management framework module that supports various microservice applications, a RAGS knowledge base retrieval module, and an agent development and deployment module for building various agent applications. (5) Service layer: The platform provides various automated business process services, knowledge base retrieval services, API interface services, database retrieval services, and other various large model training services based on intelligent question answering. (6) Application layer: The current application scenarios implemented by the platform include intelligent question answering, policy document summarization, statistical report generation, AI automatic business processes, intelligent analysis, etc. (7) Visual layer: The platform supports the function of converting analysis data into analysis reports and can build various analysis report visualizations for users with different roles and permissions.

[0072] Preferably, in some embodiments, the SQL processing module includes:

[0073] A requirements analysis unit for determining the target database type and the type of SQL statements to be supported, as well as the conversion requirements from natural language to SQL;

[0074] A data collection unit for collecting database SQL statement samples and natural language query samples and their corresponding SQL statements;

[0075] A data preprocessing unit for standardizing the collected SQL statements and performing natural language processing operations on natural language queries, including word segmentation and part-of-speech tagging;

[0076] A model selection and training unit for selecting a suitable machine learning or deep learning model and using the preprocessed data to train the model so that it can understand natural language and generate corresponding SQL statements;

[0077] A model optimization unit for evaluating the model, optimizing the model according to the evaluation results, and integrating the trained model into the agent platform.

[0078] As Figure 3 shown, the specific construction process of the SQL processing module 110 includes the following steps:

[0079] S111: Requirement Analysis:

[0080] Determine the target database type (such as MySQL, PostgreSQL, Oracle, etc.); determine the types of SQL statements to be supported (such as SELECT, INSERT, UPDATE, DELETE, etc.); determine the conversion requirements from natural language to SQL, including complexity, flexibility, etc.

[0081] S112: Data Collection

[0082] Collect SQL statement samples of each mainstream database; collect natural language query samples and their corresponding SQL statements.

[0083] S113: Data Preprocessing

[0084] Perform standardization processing on the collected SQL statements to ensure uniform format; perform natural language processing operations such as word segmentation and part-of-speech tagging on natural language queries.

[0085] S114: Model Selection and Training

[0086] Select a suitable machine learning or deep learning model, such as Seq2Seq, Transformer, etc.; use the preprocessed data to train the model so that it can understand natural language and generate corresponding SQL statements.

[0087] S115: Model Evaluation and Optimization

[0088] Evaluate the model and check its conversion effect on different types of SQL statements; optimize the model according to the evaluation results to improve the accuracy and efficiency of conversion.

[0089] S116: Integration and Testing

[0090] Integrate the trained model into the platform; test the platform to ensure that the natural language to SQL conversion function works properly.

[0091] S117: User Interface Design

[0092] Design a user-friendly interface that allows users to input natural language queries; design a result display interface to show the converted SQL statements.

[0093] S118: Deployment and Maintenance

[0094] Deploy the platform to the production environment; regularly maintain and update the model to adapt to new SQL statements and natural language queries.

[0095] The platform constructed through the above steps needs to sort out the SQL of existing mainstream databases, and through continuous training of SQL statements, enable the platform to understand the meanings of various SQL statements, help the platform better convert various natural languages into SQL statements, and then facilitate college students to query grades, course schedules, library collections, etc., improving the query efficiency.

[0096] Furthermore, in some embodiments, the SQL processing module further includes an agent interaction layer. The agent interaction layer uses a Bi-LSTM+CRF hybrid model to implement business intent recognition, constructs a dynamic business topology graph, dynamically adjusts node weights according to real-time access volume, provides a recommended entry and a confidence score for users, and supports multi-round dialogue management, and can perform context understanding according to user feedback.

[0097] Among them, constructing a dynamic business topology graph includes: graph construction: constructing an initial business topology graph according to the business process and the dependency relationship between nodes. Among them, nodes represent business steps or services, and edges represent the transfer relationship between steps. Dynamic adjustment: Real-time monitor the access volume of each node, and dynamically adjust the node weights according to the access volume. The weights reflect the importance or popularity of the nodes and are used for sorting the recommended entries. Confidence score: Calculate a confidence score for each recommended entry, based on historical data, user behavior, and model prediction. The confidence score is used to guide the user to select the most suitable business entry.

[0098] In this embodiment, by dynamically adjusting the business topology graph to adapt to real-time demand changes and introducing a confidence score, the recommendation accuracy can be improved. For example, during the peak period, the system automatically increases the weight of the "query account balance" node and gives it priority display in the recommended entries.

[0099] Among them, multi-round dialogue management includes: context understanding: by maintaining the dialogue state, tracking the information and intentions of the user in the previous round of dialogue; using the output of the Bi-LSTM+CRF model and the dialogue history for context understanding. Dialogue process management: According to the user feedback and the current dialogue state, determine the next dialogue process or business step, support dialogue interruption, backtracking, and jumping, improving the flexibility and robustness of the dialogue.

[0100] In this embodiment, by combining context information and user feedback, the coherence and accuracy of multi-round dialogue are realized, complex dialogue process management is supported, and user satisfaction can be improved. For example, User: "What should I do if I forget my password?" System: "You can reset your password through the mobile phone verification code. What is your mobile phone number?" After the user provides the mobile phone number, the system continues to guide the user to complete the password reset process.

[0101] Among them, the Bi-LSTM+CRF hybrid model is constructed by combining a bidirectional long short-term memory network (Bi-LSTM) and a conditional random field (CRF). Bi-LSTM (bidirectional long short-term memory network): It is used to capture the temporal dependencies in the user input statement, processes sequence data simultaneously from both the forward and backward directions, and can better understand the context. CRF (conditional random field): Based on the output of Bi-LSTM, it further considers the dependencies between labels, globally optimizes the sequence labeling problem, and improves the accuracy of intent recognition. The specific components of this hybrid model include:

[0102] An input layer, which is used to receive text data and convert each character or word into a word vector representation;

[0103] A word embedding layer, which is used to map the word vectors into a high-dimensional space to better capture the semantic relationships between words;

[0104] The Bi-LSTM layer, which consists of a forward LSTM and a backward LSTM, can capture both the forward and backward information in the sequence simultaneously, and the output at each time step contains the context information of the current position. Its output is the hidden state of each word;

[0105] A fully connected layer, which is used to map the output of the Bi-LSTM layer into the label space and output the scores for each position corresponding to each label;

[0106] The CRF layer, which is used to globally optimize the output of the fully connected layer according to the dependencies between labels. The output of Bi-LSTM is used as the input of CRF, and by calculating the transition matrix and the emission matrix, it outputs the optimal label sequence (i.e., the business intent).

[0107] The process of implementing business intent recognition using the above Bi-LSTM+CRF hybrid model mainly includes the following steps:

[0108] (1) Forward propagation:

[0109] Convert the input sequence into a sequence of word vectors through the word embedding layer;

[0110] Input the sequence of word vectors into the Bi-LSTM layer to obtain the hidden state at each time step;

[0111] Input the hidden state into the fully connected layer to obtain the scores for each position corresponding to each label;

[0112] Use the CRF layer to consider the dependencies between labels and calculate the optimal label sequence.

[0113] (2) Define the loss function:

[0114] Use the negative log-likelihood loss function to calculate the difference between the predicted label sequence and the true label sequence. The negative log-likelihood loss function used is:

[0115]

[0116] Where y is the true label sequence, x is the input sequence (the hidden state sequence after processing by the word embedding layer and the Bi-LSTM layer), and P(y|x) is the probability that the output sequence is y given the input sequence x.

[0117] The loss function consists of two parts: the emission score and the transfer score, which correspond to the output of the fully connected layer and the transfer matrix of the CRF layer respectively. It can be calculated by the transfer matrix and emission matrix of the CRF layer:

[0118]

[0119] Here, y′ is one of all possible label sequences. Given an input sequence x, the model may output multiple different label sequences, and y′ is one of these possible label sequences. When calculating the loss function, it is necessary to traverse all possible label sequences y′ to calculate the probability of the output sequence being y′ given the input sequence x and normalize these probabilities. Represents the set of all possible label sequences given the input sequence x; Represents all possible label sequences y′ (belonging to the set ) calculates the exponential sum of its scores (the scores obtained by the transfer matrix and emission matrix of the CRF layer).

[0120] (3) Back propagation:

[0121] By calculating the gradient of the loss function, the model parameters are updated, including the parameters of the word embedding layer, Bi-LSTM layer, fully connected layer, and CRF layer.

[0122] (4) Decoding:

[0123] During the testing or inference phase, the Viterbi algorithm is used to decode the optimal label sequence on the CRF layer.

[0124] set up is the tag sequence to be decoded, is the corresponding input sequence (the hidden state sequence after being processed by the word embedding layer and the Bi-LSTM layer).

[0125] The traditional CRF decoding formula is expressed as:

[0126]

[0127] Among them, A is the transition matrix and P is the emission matrix.

[0128] The present invention improves the traditional CRF decoding formula, introduces dynamic weight adjustment based on context information, and defines a weight adjustment factor , which dynamically adjusts the weight of the label according to the context information at the current decoding position i. Therefore, the improved decoding formula is expressed as:

[0129]

[0130] Among them, the weight adjustment factor is calculated based on the Bi-LSTM hidden state at the current position and the output of the fully connected layer. For example, in some optional embodiments of the present invention, a simple feedforward neural network (FFNN) is used to calculate based on and :

[0131]

[0132] Among them, is an activation function (such as the sigmoid function), which is used to map the output to the interval (0, 1) and thus serves as the weight adjustment factor.

[0133] During the training process, the present invention needs to optimize the parameters of the Bi-LSTM layer, the fully connected layer, and the weight adjustment factor calculation network (such as FFNN) simultaneously. This can be achieved through the backpropagation algorithm, where the loss function is still the negative log-likelihood loss function, but the influence of the weight adjustment factor needs to be considered when calculating the gradient. Specifically, the training process of the Bi-LSTM+CRF hybrid model includes: Data preparation: Collect and label the training data (including user queries, business intent labels, etc.), and convert each character or word into a word vector representation. Model initialization: Initialize the model parameters, including the parameters of the word embedding layer, the Bi-LSTM layer, the fully connected layer, and the CRF layer. Forward propagation: Input the training data into the model and perform forward propagation to calculate the loss function. Backward propagation: Calculate the gradient of the loss function and update the model parameters. Iterative training: Repeat the forward propagation and backward propagation processes until the predetermined number of training epochs is reached or the model performance no longer improves significantly. Model evaluation: Evaluate the model performance on the validation set or test set, such as accuracy, recall, F1 value, etc. Model tuning: Adjust the model parameters or network structure according to the evaluation results to improve the model performance. Through the above steps, a Bi-LSTM+CRF hybrid model with excellent performance can be trained to solve the sequence labeling problem. In practical applications, this model can be widely used in the field of natural language processing, such as named entity recognition, part-of-speech tagging, text classification, etc.

[0134] By introducing a dynamic weight adjustment mechanism based on context information, the CRF decoding process can better adapt to a specific context environment, thereby improving the accuracy and robustness of label sequence decoding. Especially when dealing with complex or ambiguous input sequences, this mechanism can significantly enhance the performance of the model.

[0135] This embodiment combines the temporal modeling ability of Bi-LSTM and the global optimization ability of CRF to improve the accuracy of intent recognition and dynamically update the model to adapt to the changes in new services or user behaviors. For example, when the user asks: "I want to query my account balance." The system uses the Bi-LSTM+CRF model to identify the business intent as "query account balance".

[0136] Furthermore, the SQL processing module further includes an NLP engine, a rule engine, and a speech engine, which are respectively used for intent recognition, rule inference, and speech processing;

[0137] Among them, the NLP engine integrates intent recognition and multi-turn dialogue management, and uses a sequence-to-sequence (Seq2Seq) model, such as Transformer or LSTM+Attention, to convert the natural language input by the user into an SQL query statement. The model selection and training unit uses the machine learning or deep learning algorithm and uses a large number of labeled natural language-SQL pairs as training data to achieve the conversion from natural language to SQL statements, supporting college users to perform data query operations;

[0138] The rule engine dynamically loads the SWRL rules provided by each business department, supporting the automated management of business processes; uses SWRL (Semantic Web Rule Language) to define business rules. The rule engine dynamically loads the SWRL rules provided by each business department, makes inferences and decisions according to the rules, dynamically loads the rules, and adapts to business changes. It supports complex business logic and rule reasoning. Application example: The business department defines a rule: "If the user's account balance is less than 100 yuan, then send a reminder notice." The rule engine loads and executes this rule. When the user's account balance is less than 100 yuan, a reminder notice is automatically sent.

[0139] The speech engine supports mixed Chinese and English speech processing, uses the Conformer model for speech recognition, and fine-tunes on the college scene corpus, while providing multiple speech synthesis modes. Among them, Conformer combines the advantages of Transformer and LSTM and has better recognition performance. Fine-tuning the model on the college scene corpus improves the recognition accuracy. Providing multiple speech synthesis modes, such as standard speech, emotional speech, etc., meets the needs in different scenarios. It supports mixed Chinese and English speech processing and adapts to multilingual scenarios. Using the advanced Conformer model improves the recognition accuracy. Providing multiple speech synthesis modes enhances the user experience. For example: The user inputs via speech: "I want to query my English score." The system recognizes and converts it into text: "I want to query my English score." The system replies in standard speech: "Your English score is 85 points."

[0140] Furthermore, the SQL processing module further includes a data service layer, which is connected to multiple business databases in real time, constructs a cross-system data federation query interface, and integrates a multi-modal feature extractor; the data service layer is connected to multiple business databases such as academic affairs, finance, and student affairs in real time, constructs a cross-system data federation query interface, and supports the storage, processing, and analysis of large-scale data. By integrating a multi-modal feature extractor, it can perform feature extraction and analysis on various types of data such as text, images, and tables, providing rich data support for upper-layer applications.

[0141] Furthermore, the vector library management module 120 includes:

[0142] A document acquisition unit, configured to acquire the API interface documents of each vector library from a preselected list of vector libraries;

[0143] A platform API interface unit, configured to provide unified API interfaces, enabling the platform to interact with different vector libraries through these interfaces, and the abstraction layer of the API interfaces can handle the differences between different vector libraries.

[0144] As Figure 4 shown, the specific construction process of the vector library management module 120 includes the following steps:

[0145] S121: Research and select vector libraries

[0146] Determine the list of vector libraries to be supported, including Milvus, Pinecone, Vespa, Weaviate, Vald, GSI, and Qdrant, etc. Research the features, performance, and API interfaces of each vector library.

[0147] S122: Acquisition of API interface documents:

[0148] Acquire the API interface documents of each vector library to understand how to use the API for data operations; for example, for Pinecone, the usage methods of API interfaces such as initialization, creating an index, retrieving the index list, and inserting data can be obtained through the official documentation.

[0149] S123: SDK development

[0150] Develop or integrate an SDK (Software Development Kit) according to the API interface documents to facilitate the use of these APIs in the platform. For example, Vald provides two types of interfaces, gRPC and REST API, and it is recommended to use gRPC for better performance.

[0151] S124: Platform API interface design

[0152] Design unified API interfaces, enabling the platform to interact with different vector libraries through these interfaces. Ensure that the abstraction layer of the API interfaces can handle the differences between different vector libraries.

[0153] S125: Platform SDK development

[0154] Develop the SDK of the platform, encapsulate the API interfaces of different vector libraries, and provide unified interfaces for developers to use. For example, the usage tutorial of Weaviate shows how to import data into Weaviate and how to perform similarity searches.

[0155] S126: Testing and verification

[0156] Test the API interfaces of each vector library to ensure that they can work properly on the platform.

[0157] Verify whether the SDK of the platform can interact correctly with each vector library.

[0158] S127: Documentation and sample code

[0159] Provide detailed API interface documentation and sample code to help developers understand and use the platform's API interfaces. For example, Milvus provides the Python SDK PyMilvus, allowing developers to interact with Milvus in a Python environment.

[0160] S128: User feedback and iteration

[0161] Collect user feedback and optimize the API interfaces and SDK according to the feedback; continuously iterate and update to support new versions and new features of the vector library.

[0162] The vector library management module 120 constructed through the above steps in this embodiment, when a certain university is managing scientific research projects and displaying academic achievements, needs to process a large amount of multi-modal data such as text, images, and audio. The vector library management module 120 can be docked and adapted with various mainstream open-source and closed-source vector libraries, support dynamic data synchronization and multi-modal joint retrieval, and can improve data processing efficiency.

[0163] Furthermore, the extensible architecture module 130 includes: a plug-in architecture unit for integrating third-party developers' custom plug-ins and providing necessary tools and libraries; a data type and index specification definition unit for defining the specifications of new data types and customized indexes; where the specifications of the new data types include data structures and operation interfaces, and the specifications of the customized indexes include index structures and index methods; a query optimization unit for implementing a query optimizer to support user-defined query optimization strategies; a plug-in management tool for facilitating users to install, configure, and uninstall plug-ins; a testing and verification unit for testing the plug-in framework and new functions and verifying whether the new data types and index types meet the data storage requirements of specific fields.

[0164] As Figure 5 shown, the specific construction process of the extensible architecture module 130 includes the following steps:

[0165] S131: Requirement analysis and planning

[0166] Analyze user requirements, determine the goals and scope of the plug-in mechanism; plan the supported data types, index types, and query optimization strategies.

[0167] S132: Design the plugin architecture:

[0168] Design a flexible plugin architecture to ensure that third - party developers can easily integrate custom plugins; define plugin interfaces and APIs to ensure compatibility with the platform.

[0169] S133: Develop the plugin framework:

[0170] Implement the plugin framework, providing the necessary tools and libraries so that developers can build plugins on this basis; including mechanisms for plugin loading, unloading, execution, and management.

[0171] S134: Define data types and indexing specifications:

[0172] Define the specifications for new data types, including data structures and operation interfaces; define the specifications for customized indexing, including index structures and indexing methods.

[0173] S135: Implement query optimization strategies:

[0174] Implement a query optimizer that supports user - defined query optimization strategies; allow users to customize query plans according to application scenarios.

[0175] S136: Develop plugin management tools:

[0176] Develop plugin management tools to facilitate users in installing, configuring, and uninstalling plugins; provide a user interface or command - line tools to simplify the management process.

[0177] S137: Testing and verification:

[0178] Test the plugin framework and new features to ensure stability and performance.

[0179] Verify whether the new data types and index types meet the data storage requirements of specific domains.

[0180] S138: Documentation and examples:

[0181] Provide detailed development documentation and sample code to help third - party developers understand how to develop plugins.

[0182] Include guidelines on how to define new data types, indexing, and query optimization strategies.

[0183] S139: Release and feedback collection:

[0184] Release the plugin mechanism and related tools; collect user feedback and perform iterative optimization based on the feedback.

[0185] S1310: Continuous support and updates:

[0186] Provide continuous technical support and updates to adapt to new technological developments and user needs.

[0187] The scalable architecture module 130 constructed through the above steps in this embodiment supports multiple models and multiple inference frameworks, and allows users to customize data types, indexing strategies, and query optimization plugins, achieving the scalability of the system and facilitating the future integration of new models, algorithms, and data sources.

[0188] Furthermore, the vertical domain expertise module 140 includes:

[0189] A data collection and preprocessing unit for collecting industry-specific datasets of higher education institutions, including text and voice data, and performing data preprocessing, including data cleaning, annotation, and structuring;

[0190] A model selection and training unit for processing materials through a multi-task learning framework, selecting appropriate machine learning or deep learning algorithms, training the model using the preprocessed industry-specific datasets, and adjusting parameters according to the model's performance to achieve multi-modal material pre-review;

[0191] A role division unit for customizing different prompt words and interaction logics according to the user's organizational structure role division, simulating the tone, intonation, and speech rate of users in the higher education industry;

[0192] A user feedback unit for collecting user feedback and performing iterative optimization based on the feedback.

[0193] Furthermore, the vertical domain expertise module 140 also includes a natural language processing model and a speech recognition model, which are respectively used to process text and voice data;

[0194] The deep learning algorithm of the speech recognition model includes a Transformer model for processing long sequence data and capturing global dependencies, and provides a rich industry knowledge base and a question-and-answer system to support users in obtaining industry knowledge and answering questions through natural language queries.

[0195] As Figure 6 shown, the specific construction process of the vertical domain expertise module 140 includes the following steps:

[0196] S141: Requirement research and analysis:

[0197] Cooperate with higher education institutions and government departments to understand their specific needs and pain points; collect industry-specific datasets, including text, voice, etc. data.

[0198] S142: Data preprocessing:

[0199] Design a data preprocessing process according to the actual database structure of the user; clean, annotate, and structure the data to meet the needs of model training.

[0200] S143: Model Design and Selection:

[0201] Design a model architecture suitable for industry characteristics.

[0202] Select appropriate machine learning or deep learning algorithms, such as natural language processing models, speech recognition models, etc. Specifically:

[0203] In view of the characteristics of the vertical field of colleges and universities, the present invention designs a hybrid model architecture that combines the advantages of deep learning and knowledge graphs. This model architecture mainly includes the following parts:

[0204] Deep learning basic model: (1) Text processing sub-model: Adopt a self-attention mechanism model based on Transformer (such as BERT or RoBERTa) to process and analyze text data. The Transformer model can capture long-distance dependencies through the self-attention mechanism, which is very suitable for processing long text data in colleges and universities, such as policy documents, academic papers, etc. (2) Speech processing sub-model: Adopt the Conformer model, which combines the advantages of Transformer and convolutional neural network (CNN), and can efficiently process speech signals, especially showing excellent performance when processing speech data in the college environment with complex background noise.

[0205] Knowledge graph fusion module: Construct a knowledge graph for the field of colleges and universities, including entities such as students, teachers, courses, scientific research projects, etc. and their relationships. Combine the knowledge graph with the deep learning model to enhance the model's understanding and reasoning ability of domain knowledge. For example, when processing a student's course selection request, the model can use the course dependency relationship (such as prerequisite courses) in the knowledge graph to provide more accurate suggestions.

[0206] After selecting the above model architecture, the present invention further selects appropriate machine learning or deep learning algorithms to train these models. The following are the specific algorithms for text and speech processing:

[0207] Specifically, for the text processing sub-model, the following text processing algorithms are used: Fine-tune on the BERT or RoBERTa basic model to adapt to the data in the college and university field. Taking the BERT basic model as an example:

[0208] Parameter definition: Word embedding dimension ( ): Dimension of word vectors in the model, usually set to 768 or 1024. Number of heads (h): Number of heads in the multi-head self-attention mechanism, usually set to 8 or 12. Dimension of the feed-forward neural network (dff): Dimension of the hidden layer in the feed-forward neural network, usually set to 3072 or 4096. Number of layers (L): Number of layers of the Transformer encoder, usually set to 12 or 24.

[0209] Input representation: Convert the input text into a sequence of word vectors, and add position encoding and segment encoding. The formula is as follows:

[0210] Input = WordEmbeddings + PositionEmbeddings + SegmentEmbeddings

[0211] Among them, WordEmbeddings is the sequence of word vectors, PositionEmbeddings is the position encoding, and SegmentEmbeddings is the segment encoding.

[0212] Transformer encoder: For each word vector, process it through the multi-head self-attention mechanism and the feed-forward neural network. The calculation formula is as follows:

[0213]

[0214] Among them, Q, K, and V are the query, key, and value matrices respectively, is the dimension of the key vector.

[0215] Feed-forward neural network:

[0216] )

[0217] Among them, and are the weight matrices, and are the bias terms.

[0218] For the speech processing sub-model, the Conformer model is used with the speech processing algorithm. Taking the self-attention mechanism in Conformer as an example:

[0219] Parameter definition: Convolution kernel size: Used for convolution operations in the convolutional layer, usually set to 3 or 5. Number of convolutional layers: Number of convolutional layers in Conformer, adjusted according to the model complexity. Number of attention heads: The same as Transformer, usually set to 8 or 12.

[0220] Position Encoding: Similar to Transformer, Conformer also uses position encoding to capture position information in the sequence, and uses sine and cosine functions for position encoding.

[0221] Self-Attention Mechanism: The same as Transformer, but Conformer adds a convolutional layer in the self-attention mechanism to capture local features. The calculation formula of the self-attention mechanism is the same as that in the above Transformer.

[0222] Convolutional Layer: After the self-attention mechanism, Conformer applies a series of depthwise separable convolutional layers to further extract features. The calculation formula of the convolutional layer is:

[0223]

[0224] where DepthwiseConv is the depthwise separable convolution, PointwiseConv is the pointwise convolution, and BN is the batch normalization.

[0225] The Knowledge Graph Fusion Module uses a Graph Neural Network (GNN) to fuse knowledge graph information. Specifically, taking the Graph Convolutional Network GCN as an example:

[0226] Parameter Definition: Learning Rate (η): The step size used for model training, adjusted according to the training situation. Weight Decay (λ): The coefficient of the regularization term used to prevent overfitting. Number of Layers (LGCN): The number of layers of GCN, usually set to 1 or 2.

[0227] Node Feature Update: For each node in the knowledge graph, its feature representation is updated by aggregating the features of its neighbor nodes. The calculation formula is as follows:

[0228] where A is the adjacency matrix, D is the degree matrix, is the node feature matrix of the l-th layer, is the weight matrix of the l-th layer, and σ is the activation function (such as ReLU).

[0229] S144: Targeted Training and Parameter Adjustment:

[0230] Use the industry user's proprietary dataset to train the model; adjust the parameters according to the model performance to optimize the model performance.

[0231] S145: Role Division and Prompt Customization:

[0232] According to the role division of the user's organizational structure, customize different prompts and interaction logics; simulate the tone, intonation, and speech rate of industry users to improve the user experience.

[0233] S146: Model Training and Validation:

[0234] Perform a large number of model trainings on the preprocessed data; verify the accuracy and adaptability of the model to ensure that it meets the needs of industry users.

[0235] S147: System Integration and Testing:

[0236] Integrate the trained model into the platform; conduct comprehensive system testing to ensure the stability and reliability of the system.

[0237] S148: User Feedback and Iterative Optimization:

[0238] Collect user feedback and evaluate the actual application effect of the model; perform iterative optimization based on the feedback to improve the accuracy of the model and the user experience.

[0239] S149: Deployment and Monitoring:

[0240] Deploy the optimized model to the production environment.

[0241] Implement monitoring to ensure the stable operation of the system and respond promptly to possible problems.

[0242] S1410: Documentation and Training:

[0243] Provide detailed system documentation and user manuals.

[0244] Train users to ensure that they can make full use of the platform functions.

[0245] In the embodiments of the present invention, the vertical domain professional knowledge module 140 is constructed through the above steps. A large amount of preprocessing and pre-training are performed on the data in the vertical domain of colleges and universities, improving the system's understanding and processing capabilities for specific business scenarios, thereby outputting material completeness scores, missing item detection, and urgency classification information, and adjusting parameter settings according to industry characteristics, and customizing and optimizing the interaction experience through role-based prompt words.

[0246] In summary, the present invention realizes the intelligent processing and analysis of multi-source data by integrating SQL processing capabilities, vector library management, scalable architecture, and vertical domain professional knowledge applications, significantly improving the business handling efficiency and user experience of the college online service hall, and contributing to the digital transformation and intelligent upgrade of the education industry.

[0247] The above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the embodiments, those of ordinary skill in the art should understand that: still, the specific implementation manners of the present invention can be modified or equivalently replaced, and any modification or equivalent replacement that does not depart from the spirit and scope of the present invention shall be covered by the scope of the claims of the present invention.

Claims

1. An AI intelligent one-stop service system for colleges and universities based on an agent platform, characterized in that, It includes: An SQL processing module, which is used to organize the SQL statements of existing mainstream databases, and convert the natural language input by users into SQL statements of the target database through pre-trained machine learning or deep learning models; Among them, the SQL processing module includes: A requirement analysis unit, which is used to determine the target database type, the type of SQL statements to be supported, and the conversion requirements from natural language to SQL; A data collection unit, which is used to collect database SQL statement samples, natural language query samples, and their corresponding SQL statements; A data preprocessing unit, which is used to perform standardization processing on the collected SQL statements, and perform natural language processing operations on natural language queries, including word segmentation and part-of-speech tagging; A model selection and training unit, which is used to select machine learning or deep learning models, and use the preprocessed data to train the models so that they can understand natural language and generate corresponding SQL statements; A model optimization unit, which is used to evaluate the models, optimize the models according to the evaluation results, and integrate the trained models into the intelligent agent platform; A vector library management module, which docks and adapts with various mainstream open-source and closed-source vector libraries, interacts with different vector libraries through a unified API interface, and supports dynamic data synchronization and multimodal joint retrieval; An extensible architecture module, which supports multiple models and multiple inference frameworks, including submitting queries, the retriever component selecting relevant information, and the re-ranker refining the selection, is used to process data from databases, web pages, or PDF files, and continuously optimizes the data-driven engine through proxy and plugin mechanisms, and ensures the scalability of the platform through the construction of a knowledge base and intelligent agent workflow components; user-defined data types, index strategies, and query optimization plugins; A vertical domain expertise module, which performs data preprocessing and pre-trained models on data in the university vertical domain, outputs material completeness scores, missing item detection, and urgency classification information, and adjusts parameter settings according to industry characteristics to customize role-based prompt words.

2. The system according to claim 1, wherein The vector library management module includes: A document acquisition unit, which is used to obtain the API interface documents of each vector library from the preselected list of vector libraries; A platform API interface unit, which is used to provide a unified API interface, enabling the platform to interact with different vector libraries through these interfaces, and the abstraction layer of the API interface is used to handle the differences between different vector libraries.

3. The system according to claim 1, wherein The extensible architecture module includes: A plugin architecture unit, which is used to integrate third-party developer custom plugins and provide tools and libraries; A data type and index specification definition unit, which is used to define the specifications of new data types and the specifications of customized indexes; among them, the specifications of the new data types include data structures and operation interfaces, and the specifications of the customized indexes include index structures and index methods; A query optimization unit, which is used to implement a query optimizer and support user-defined query optimization strategies; A plugin management tool, which is used to facilitate the installation, configuration, and uninstallation of plugins; A testing and verification unit, which is used to test the plugin framework and new functions, and verify whether the new data types and index types meet the data storage requirements.

4. The system according to claim 1, characterized in that, The vertical domain expertise module includes: A data collection and preprocessing unit for collecting university industry-specific datasets including text and voice data, and performing data preprocessing, including data cleaning, annotation, and structuring; A model selection and training unit for processing materials through a multi-task learning framework, selecting machine learning or deep learning algorithms to build a model, training the model using the preprocessed industry-specific dataset, and adjusting parameters according to the model performance to achieve multi-modal material pre-review; A role division unit for customizing different prompt words and interaction logics according to the user organization structure role division, and simulating the tone, intonation, and speech rate of university industry users; A user feedback unit for collecting user feedback and performing iterative optimization based on the feedback.

5. The system according to claim 1, wherein The SQL processing module further includes an agent interaction layer. The agent interaction layer uses a Bi-LSTM+CRF hybrid model to implement business intent recognition, constructs a dynamic business topology map, dynamically adjusts node weights according to real-time access volume, provides recommended entrances and confidence scores for users, and supports multi-turn dialogue management, and can perform context understanding based on user feedback.

6. The system according to claim 5, characterized in that, Among them, The Bi-LSTM+CRF hybrid model is constructed by combining a bidirectional long short-term memory network and a conditional random field, including: An input layer for receiving text data and converting each character or word into a word vector representation; A word embedding layer for mapping the word vector to a high-dimensional space to capture the semantic relationships between words; A Bi-LSTM layer composed of a forward LSTM and a backward LSTM, which can capture both forward and backward information in the sequence and the output of each time step contains the context information of the current position; A fully connected layer for mapping the output of the Bi-LSTM layer to the label space and outputting the scores of each position corresponding to each label; A CRF layer for globally optimizing the output of the fully connected layer according to the dependency relationships between labels.

7. The system according to claim 5, wherein Among them, The SQL processing module further includes an NLP engine, a rule engine, and a voice engine, which are used for intent recognition, rule inference, and voice processing respectively; Among them, the NLP engine integrates intent recognition and multi-turn dialogue management. The model selection and training unit trains SQL statements through the machine learning or deep learning algorithm to achieve the conversion from natural language to SQL statements and supports university users to perform data query operations; The rule engine dynamically loads SWRL rules provided by each business department and supports the automated management of business processes; The voice engine supports mixed Chinese and English voice processing, uses a Conformer model for speech recognition, and is fine-tuned on university scenario corpora, and at the same time provides multiple voice synthesis modes.

8. The system according to claim 6, wherein The SQL processing module further includes a data service layer, which is connected to multiple business databases in real time, constructs a cross-system data federation query interface, and integrates a multi-modal feature extractor; the data service layer is connected to multiple business databases such as academic affairs, finance, and student affairs in real time, constructs a cross-system data federation query interface, and supports the storage, processing, and analysis of large-scale data. By integrating a multi-modal feature extractor, it can perform feature extraction and analysis on various types of data such as text, images, and tables, providing data support for upper-layer applications.

9. The system according to claim 4, characterized in that The vertical domain expertise module further includes a natural language processing model and a speech recognition model, which are respectively used to process text and speech data. The deep learning algorithm of the speech recognition model includes a Transformer model for processing long sequence data and capturing global dependencies, and provides an industry knowledge base and a question-and-answer system to support users in obtaining industry knowledge and answering questions through natural language queries.

Images